R help - Jan 2006 - non-finite finite-difference value[]

Dear R-helpers,

running a zeroinflated model of the following type:
zinb = zeroinfl(count=response ~., x = ~ . - response, z = ~. - response,
dist = "negbin", data = t.data, trace = TRUE)
generates the following message:

Zero-Inflated Count Model
Using logit to model zero vs non-zero
Using Negative Binomial for counts
dependent variable y:
Y
  0   1   2   3
359  52   7   3
generating start values...done
MLE begins...
initial  value 262.883959
error in optim(fn = llhfunc, par = stval, method = method, control control,  :
        non-finite finite-difference value [3]


Short version of my question:
what information does non-finite finite-difference value [3] give?


Extended version of my question:
Since zeroinfl() calls optim(), I assume, that an infinite value is
generated during the iteration steps of the optimizing algorithm. So the
value[3] gives me a hint to the step, when it happens?
Or expressed in a different way, are my data structured in a way, that
finding the maximum in the response surface is highly sensitive of the
initial value, used for the algorithm?
And if this is correct, is there an easy way in R, to set another starting
value that the one, which is generated automatically in zeroinfl() (I mean,
just from the technical point of view, without considering the theoretical
thrill, which will be my next problem, since it is not at all familiar to
me.).
Or last questions: how can I extract the function, generated by zeroinfl()
and passed to optim()? Is there a possibility to visualize it, so that I get
an idea, what is going on?

Any hint, which will help me to dig further in the right direction is more
than welcome.
Thank you in advance.
Kind regards,
P. Olsson


PS: Sorry, I tried, but I did not manage to set up an example which
generates the above mentioned message.
PPS: I am running R 2.2.1 under windows.

	[[alternative HTML version deleted]]

On Tue, 24 Jan 2006, P. Olsson wrote:
> Dear R-helpers,
>
> running a zeroinflated model of the following type:
> zinb = zeroinfl(count=response ~., x = ~ . - response, z = ~. - response,
> dist = "negbin", data = t.data, trace = TRUE)
> generates the following message:
>
> Zero-Inflated Count Model
> Using logit to model zero vs non-zero
> Using Negative Binomial for counts
> dependent variable y:
> Y
>  0   1   2   3
> 359  52   7   3
> generating start values...done
> MLE begins...
> initial  value 262.883959
> error in optim(fn = llhfunc, par = stval, method = method, control >
control,  :
>        non-finite finite-difference value [3]
>
>
> Short version of my question:
> what information does non-finite finite-difference value [3] give?
>
>
> Extended version of my question:
> Since zeroinfl() calls optim(), I assume, that an infinite value is
> generated during the iteration steps of the optimizing algorithm. So the
> value[3] gives me a hint to the step, when it happens?
No.  The gradient is a vector, which is (I assume from the optim call) 
being evaluated by finite differences.  It is the third element of the 
vector that is non-finite (infinite or NA or NaN).

The information is for the optim user, which in this case is zeroinfl,
and is probably best passed on to the author of that function with an 
example.
> Or expressed in a different way, are my data structured in a way, that
> finding the maximum in the response surface is highly sensitive of the
> initial value, used for the algorithm?
> And if this is correct, is there an easy way in R, to set another starting
> value that the one, which is generated automatically in zeroinfl() (I mean,
> just from the technical point of view, without considering the theoretical
> thrill, which will be my next problem, since it is not at all familiar to
> me.).
> Or last questions: how can I extract the function, generated by zeroinfl()
> and passed to optim()? Is there a possibility to visualize it, so that I
get
> an idea, what is going on?
These are questions about zeroinfl.  That is not part of R, and we don't 
know what it is.  There is a function of that name in package pscl, 
but you have not mentioned that and there could be others.

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595