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

[dave fournier]

Anoither way to  do this is to add a tiny quadratic penalty to the
original model

SEPARABLE_FUNCTION void f1(int& i,const dvar_vector& log_tau,const 
prevariable & rho,const dvar_vector & bbi,const dvar_vector & beta)
   dvar_vector tau=exp(log_tau);
   if (i==1)
   {
     ll+=1.e-6*norm2(log_tau);
     ll+=1.e-6*square(rho);
   }


You may need to use something like

    -crit 1.e-8


I have done enough simulations to see a pattern.

Laplace approx

ADMB-RE   0.12426 169 0.973034 0.159918 0.141409
lmer           0.252941 170 0.985313 0.200662 0.118773

so the type 1 error rates are approx 0.12 and 0.25

and for 5 point AGHQ

ADMB-RE  0.0666667 105 0.986444 0.129637 0.140441
lmer          0.333333 105 0.872585 0.167648 0.107097


Of course these samples are a bit small to say anything definitive.

It would be good to modify the R code so that the random numbers produced
can be controlled so that we all get the same results for the same seed.