[ADMB Users] Restricting magnitude of random effects estimates, achieving convergence of RE models

[Chris Gast]

Thanks for the explanation, Hans.  I still haven't experimented with the
effect of -noinit on convergence, as I've been toying with the
logistic-transformed "bounded" random effects you suggested earlier.

I'm not having much luck getting them to work, and I think there's a valid
reason for it.

I define

init_bounded_number logtau(-10,3,2);
random_effects_vector el(0,n-2,2);
vector e(0,n-2);

in the parameter section (among many other things, of coures), and then at
the top of the procedure section, I compute

tau=mfexp(logtau);

for(i=0;i<(n-1);i++){
e[i] = -5 + 10*exp(el[i])/(1+exp(el[i]));
}

and then proceed to use el[i] values within the model, including

totL += -(n-1)*log(tau)-.5*norm2(e/tau);

with totL the objective function value.  These random effects occur within a
logistic function themselves [survival probability =
exp(s+e[i])/(1+exp(s+e[i]))], so an upper limit on tau of exp(3) should
quite easily cover the range of survival probability within 0 and 1 along
with e[i] ranging from about -5 to 5.  Also, my simulated datasets do not
contain an inordinate amount of variation in this survival probability, but
enough such that detection should be relatively easy.

I believe the problem is that, in this case, an el[i] value of 6 is just as
good as an el[i] value of 60 (they lead to the same result, namely that
logit(el[i])=1).  Thus, there are many, many el[i]'s to get the same e[i],
and I end up with random effects values steadily increasing in magnitude
into the thousands, until I finally cease optimization.

Just wanted to update anyone who might be following along with my progress.
 I'll keep working on a solution, but still welcome any other suggestions
that might be out there.



Chris








-----------------------------
Chris Gast
cmgast at gmail.com


On Wed, Aug 4, 2010 at 10:16 PM, H. Skaug <hskaug at gmail.com> wrote:

> Chris and Mark,
>
> No, by default the RE are initialized to zero (or whatever is found in
> the .pin file)
> at the start of each inner optimization (part of the Laplace
> approximation).
> This has been found to be more robust in general than using the RE estimate
> from the previous Laplace appr.
>
> I hoped that it could help here, but it is a shot in the dark. (One
> could also hope that -noinit reduces the computational time somewhat)
>
> Hans
>
>
> > I suppose I'm a bit confused about the -noinit option.  Shouldn't ADMB
> use
> > the previous RE estimates as the starting point for the next optimization
> by
> > default?  Perhaps I'm misunderstanding something.
> > Thanks again,
> > Christ
> >
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/users/attachments/20100805/5b0ef3d7/attachment.html>