[R-sig-ME] Negative Binomial in glmmadmb

[Ben Bolker]

Chad Newbolt <newboch at ...> writes:

>

[snip] 

> Since I have evidence for overdispersion, I'm using negative
> binomial distribution as opposed to Poisson.  My two questions are:
> 1) When I fit using the following global zero inflation model I
> receive the following error:
 
> fit1=glmmadmb(Fawn~Age+I(Age^2)+BodySize+SSCM+AvgAge+Age*AvgAge+
  I(Age^2)*AvgAge+BodySize*AvgAge+SSCM*AvgAge+(1|Sire),
   data=datum,family="nbinom",zeroInflation = TRUE)

I think you can shorten this a bit to

(Age+I(Age^2)+BodySize+SSCM)*AvgAge + (1|Sire)

> Parameters were estimated, but standard errors were not: the most
> likely problem is that the curvature at MLE was zero or negative
> Error in glmmadmb(Fawn ~ Age + I(Age^2) + BodySize + SSCM + AvgAge +
> Age * : The function maximizer failed (couldn't find parameter file)
> Troubleshooting steps include (1) run with 'save.dir' set and
> inspect output files; (2) change run parameters: see
> '?admbControl';(3) re-run with debug=TRUE for more information on
> failure mode In addition: Warning message: running command
> 'C:\windows\system32\cmd.exe /c glmmadmb -maxfn 500 -maxph 5 -noinit
> -shess' had status 1
 
> However, when I change to zeroInflation = FALSE, I receive no
>  warnings and everything seems to go as should.
 
> Does this simply mean that my data is not zero inflated, hence the
> zero inflated model will not run, or is this something I should be
> concerned about and investigate the cause further?  When I debug I
> see the following warning....Warning -- Hessian does not appear to
> be positive definite Hessian does not appear to be positive
> definite.

> 2) When fitting more simple versions(predictors removed) I receive
> the same error as above when using the family=nbinom; however these
> errors disappear when using family=nbinom1.  Is this indicative of
> an underlying problem or am I OK to use the ouput from the later
> family where variance = ??.  Thanks, Chad [[alternative HTML version
> deleted]]

  Short answer: you should be a little concerned, and you should
not assume that your data are not zero-inflated. These are not
indications about what your model is actually finding, just indications
that ADMB ran into *some* kind of trouble. Unfortunately,
there is no really simple guide to trouble-shooting these kinds of
problems.  Some general suggestions:

* try out the glmmTMB package - it's newer/experimental, but
often more stable
* the ?admbControl man page suggests trying shess=FALSE and noinit=FALSE
* it may not help in this case, but centering continuous predictors is 
always worth a shot
* similarly, poly(Age,2) is a little more stable than (Age+I(Age^2))
* inspect your data graphically to see whether there are outliers 
or other odd patterns that might be messing up the fit