[ADMB Users] SE Estimates including process error in ADMB

dave fournier otter at otter-rsch.com
Fri Mar 18 09:10:49 PDT 2011

I think the right way is to bound the normal after.  Say  u is a RE.
the - log prior is  u^2/2

You define a transformed u say call it v where

         v = f(u)

and f(u) = u for -a<u<a
and  f(u) = a + (u-a)/(1+(u-a)) for a<u
similar for u<-a
and use  sigma*v as the RE in the model.  Proably a should have a value 
like 3.  When convergence is achieved you can then probably get rid of f 
and just use u.  Note f is not quite smooth enough for the theory, but 
it proabably works for an initial fitting.

More information about the Users mailing list