[ADMB Users] quick, stupid question: retrieve standard errors for some parameters when bad hessian?

[dave fournier]

On 11-11-02 09:25 AM, Ben Bolker wrote:

That is the least interesting case. However what one can do is to 
ocnvince oneself that
when the model wants to set the variance equal to 0 you get the same 
estimates by
fixinf the variance at a small number (and fixing th covariance too).

   init_vector log_tau(1,2,-1)
   init_bounded_number rho(-3.,3.,-1)

then make a pin file with the same values as in the ifnal estimates
and run the model

# log_tau:
  -6.29298 -5.67007
# rho:
0.00000000000
# beta:
   0 0
# bb:
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0
   0 0 0 0 0 0 0 0 0 0

and now you get the std devs for beta

index   name   value      std dev
      1   beta   9.9124e-01 2.1486e-01
      2   beta   1.1868e+00 2.4379e-01

So how to convince yourself that these are the same
std devs that you would get for the original model.

You cad add penalties on the parameters

    f+=100*square(log_tau(1)+6.2929)

etc in the original model.

this adds a term to the Hessian which will make it positive definite.
You will see that it does not change the estimate for the std dev of the 
betas.

The point is that the estimates for beta and log_tau are uncorrelated when
tau is (almost) 0.


>     In the continuing Zhang et al 2011 saga:  one of the conditions under
> which Zhang et al did the simulations was tau=0.001 (i.e. very low
> among-individual variation) -- i.e. the problem really reverts to a
> non-mixed-model problem.
>
>    When I try to do this one, ADMB perfectly reasonably complains that
> the Hessian is not positive definite.  Therefore, the std file doesn't
> get created.  I'd like to try to retrieve the Hessian and get the
> inverse of specified columns myself (or do a generalized inverse). I see
> that there are files produced:
>
>    binary: admodel.hes, analyzer.rhes, hesscheck, hessian.bin
>
>    are these documented somewhere, or has someone figured out what they
> are, or do I need to work through an example myself to figure it out ... ?
>
>    cheers
>      Ben Bolker
>