[Developers] Random effects without penalties work: should they?

Hans J. Skaug Hans.Skaug at math.uib.no
Fri Dec 14 11:45:40 PST 2012


Hi,

I verified that your explanation is correct. Then this is understood
behaviour, although not uniform across model types.

Thanks for the explanation.

Hans

From: dave fournier [mailto:davef at otter-rsch.com]
Sent: Friday, December 14, 2012 7:49 PM
To: developers at admb-project.org
Cc: Hans J. Skaug
Subject: Re: [Developers] Random effects without penalties work: should they?

On 12-12-14 09:14 AM, Hans J. Skaug wrote:

Probably because the Hessian is block diagonal. the code does the minimization  separately
for each separable call. So variables that are never used do not appear in any separable call
and never enter into the estimation.  If you used one of the RE's in two separable calls it would
probably crash.


Hi,



I have noted that if you define random effects in your model,

but do not use them for anything [u(M+1) and u(M+2)) in

the attach model] the program runs fine. This

used to cause problems, because the Laplace approximation

is not well defined.



My guess is that the reason why this now works is that

the effort to avoid the "infinite loop" in the inner problem

that was carried out some years ago has introduced

some penalty on the random effects.



Maybe the new behaviour is OK, but it means that you are no longer

told if you introduce a random effect that is not used for anything.

(Probably you did not do this unintentionally.



Mostly I was curious about finding out what has caused the change.



Hans




_______________________________________________

Developers mailing list

Developers at admb-project.org<mailto:Developers at admb-project.org>

http://lists.admb-project.org/mailman/listinfo/developers

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.admb-project.org/pipermail/developers/attachments/20121214/6086abbe/attachment-0001.html>


More information about the Developers mailing list