[ADMB Users] sdreport mystery

[Paul.Conn at noaa.gov]

Hi Rick (and others),

Thanks for the responses!  I am getting reasonable point estimates of MSY parameters, so that is why I was so confused.

Looking back at the MSY part of our code, I think I'm seeing the reason why this is happening - the function we're using to calculate MSY uses a fine spaced grid to calculate SY for different levels of F and then returns the maximum - so F/FMSY isn't continuous w.r.t. the other parameters in the model - instead it's a step function.  So it appears I'll need to try replacing that with an internal continuous maximization algorithm, if possible.  

All this comes up because we had been using bootstrapping to get at variance but were interested in how the asymptotic variance from the Hessian compared.

Cheers (and thanks!),
Paul


----- Original Message -----
From: Richard Methot <Richard.Methot at noaa.gov>
Date: Wednesday, February 10, 2010 8:00 am
Subject: Re: [ADMB Users] sdreport mystery
To: Paul Conn <Paul.Conn at noaa.gov>
Cc: users at admb-project.org


> You've got the basics of ADMB close enough.  I suggest checking the 
>  output of your sdreport quantities after running the model with the 
>  -nohess option so see if they are getting sensible quantities.
>  
>  Paul Conn wrote:
>  > Hi all:
>  > I have an sdreport mystery having to do with a statistical 
> catch-age 
>  > model.  Let me try to set the mood:
>  >
>  > It's a dark and stormy night and I'm at my computer.  I try the 
>  > following:
>  > Run 1:  A model with F/Fmsy and S/Smsy declared as a 'vector'.  The 
> 
>  > Hessian converges properly and I am exceedingly happy. Run 2:  A 
> model 
>  > with F/Fmsy and S/Smsy declared as 'sdreport_vector'.  The Hessian 
> 
>  > appears to converge but then there's a bunch of extra DOS output 
>  > saying that the estimated correlation matrix isn't positive 
> definite.  
>  > (this also happens if I declare terminal stock status as a 
>  > sdreport_number).
>  >
>  > So I'm wondering if anyone has insight on why this is happening.  
> What 
>  > I would guess ADMB is doing in the background is calculating 
> numerical 
>  > derivatives of the sd_report variables w.r.t. the actual parameters 
> 
>  > and then using a Taylor series (delta method) approximation to get 
> a 
>  > variance estimate.  If this is the case, is there a way to check 
> these 
>  > numerical derivatives to pinpoint the problem?  Or am I completely 
> 
>  > wrong about what ADMB is doing here?
>  >
>  > Thanks in advance for any thoughts!
>  >
>  > Cheers, Paul
>  >
>  _______________________________________________
>  Users mailing list
>  Users at admb-project.org
>