[ADMB Users] Discrepancy between MLEs in the .std and .rep files

[r.millar at auckland.ac.nz]

Dear ADMB'ers,

First, I'd like to give my thanks to the ADMB foundation team for making
ADMB freeware. Also, I'm really appreciative to Jon Schnute for the
PBSadmb package and the excellent installation instructions. I gave them
to a grad student a few days ago, and he was able to get ADMB up and
running without any further input from me (a first!).

In fact, I've been enjoying ADMB-RE so much that I might have come across
a little glitch. It is that the .std and .rep file will return slightly
different values for the std error of a random effect vector. For example,
in the orange tree example (obtained from ADMB webpages), I added a few
lines to the REPORT section. The .std file looks like

index   name         value      std dev
     1   beta         5.3157e-02 1.5658e+01
     2   beta         1.9063e+00 3.5249e+01
     3   beta        -7.9269e+00 2.7080e+01
     4   log_sigma    2.0596e+00 1.2910e-01
     5   log_sigma_u  3.4561e+00 3.2412e-01
     6   u           -2.9564e+01 1.4740e+01
     7   u            3.1730e+01 1.4742e+01
     8   u           -3.7196e+01 1.4759e+01
     9   u            4.0227e+01 1.4768e+01
    10   u           -5.1974e+00 1.4700e+01

and the .rep file produces (with my obvious addition of code)

log_sigma = 2.05962
log_sigma_u = 3.45462
sigma = 7.84301
sigma_u = 31.6463

Note that log_sigma_u's are different between the two files. The
difference is not much here, but it is clearly not round-off error. I had
earlier been running a linear normal one-way RE model and the relative
difference was much greater. I also had the MLE from R function lmer, and
can say that unfortunately it is the .std file that is in error (as is the
.cor file). Also, I've found that changing the bounds on the parameter
specification (but keeping them inactive) of log_sigma_u altered its value
in the .std file, but not the .rep file.

Any ideas?

Regards,

Russell