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

[dave fournier]

   This turns out to be remarkably difficult and boring to deal with.

  The idea is that first you fit the model and then check for very small 
variances

   tau1 and tau2.  If one is small you make it inactive and then 
calculate the
Hessian for the rest. Easiest is to run the model twice. Before running 
it the first time
get rid of analyzer.par

rm analyzer.par
./analyzer -ind analyzer.dat -mno 100000 -gbs 200000000 -cbs 100000000 
-nohess
./analyzer -ind analyzer.dat -mno 100000 -gbs 200000000 -cbs 100000000 
-phase 10 -ainp analyzer.par
   grep "beta2" *.std >> b1

This extra code is added to the model to determine what variables to 
turn off

DATA_SECTION
   init_int n
   int lflag1
   int lflag2
   init_matrix x(1,n,1,3)
   init_matrix Y(1,n,1,3)
  LOC_CALCS
    lflag1=2;
    lflag2=2;
    double lt1,lt2;
    lt1=0;
    lt2=0;
   cifstream cif("analyzer.par");
   if (!(!cif))
   {
     cif >> lt1 >> lt2;
     if (lt1<-3.9)
       lflag1=-1;
     if (lt2<-3.9)
       lflag2=-1;
   }


and

  init_bounded_number log_tau1(-4.0,4.0,lflag1)
  init_bounded_number log_tau2(-4.0,4.0,lflag2)

I attached the tpl file for the modified model.
-------------- next part --------------
An embedded and charset-unspecified text was scrubbed...
Name: analyzer.tpl
URL: <http://lists.admb-project.org/pipermail/users/attachments/20111103/fea13ccf/attachment.ksh>