[ADMB Users] Reliability of Hessian Matrix Values....

[Ian Taylor]

Hi Mike,
There's a little discussion on this topic from this email list archived 
here: 
http://groups.google.com/group/admb-users/browse_thread/thread/f6cc1c78acebdb2d/f11160addca05952?lnk=gst&q=hessian#f11160addca05952
I'm sure others may be able to provide more detailed support of this 
conclusion, but my experience, based on comparison with MCMC estimates 
and stability across multiple runs of similar models, is that the 
Hessian matrix values are extremely reliable, regardless of model 
complexity.
-Ian

UW_Grad_2010 wrote:
>  Hi my name is mike,
>  I am a graduate student at the University of Washington and I have
> just
>  begun using AD Model Builder. I have some questions about the
> reliability of standard
>  errors output by ADMB.
>
>  Knowing that the Standard errors are calculated from the Hessian
> matrix, and assuming that ADMB converts the Hessian matrix values to
> standard errors correctly
>
>  It then stands to reason that my question is really about the
> reliability of the Hessian Matrix values that
>  are estimated by ADMB.
>
>  Recently it has come to the attention of our working group that
> hessian matrix values from solely
>  numerically estimated models, such as in R or other programs using
> standard optimization routines, are variable based on optimizer type
> and step size. Making them unreliable.
>
>  Since ADMB has the Auto Diff feature we were hoping that this would
> make
>  the SE's estimated from the Hessian matrix stable regardless of
> optimizer
>  and step size.
>
> Does anyone have any information about what I have said here
>
> or what they find to be a suitable answer to the question:
>
> How reliable are the standard error estimates (and so in turn the
> hessian matrix values) estimated by ADMB for models with a large
> number of parameters?
>
> Thanks
> Mike
> _______________________________________________
> Users mailing list
> Users at admb-project.org
> http://lists.admb-project.org/mailman/listinfo/users
>