[ADMB Users] difference between ADMB-RE and R/mgcv in SEs for smoother coefficients in a GAM fitted by maximum likelihood

[dave fournier]

A very simple example shows how the method works.


DATA_SECTION

PARAMETER_SECTION
   init_number x
   random_effects_vector u(1,1)
   objective_function_value f
   sdreport_number sd1
   sdreport_number sd2
PROCEDURE_SECTION

   f= square(x-10) + 0.5*square(u(1)) + square(u(1)-x);
   sd1=u(1)-x;
   sd2=x-u(1);


The logarithm of the determinant of the hessian = 0.980829
  index   name    value      std.dev       1       2       3
      1   x    7.5002e+00 6.1237e-01   1.0000
      2   u    5.0001e+00 5.7735e-01   0.0000  1.0000
      3   sd1 -2.5001e+00 8.4163e-01  -0.7276  0.6860  1.0000
      4   sd2  2.5001e+00 8.4163e-01   0.7276 -0.6860 -1.0000  1.0000

I guess the question is whether this is reasonable for sd1. It seems a 
bit large to me.