[ADMB Users] Another sdreport mystery

[Shanae Allen]

Hello ADMB folks!

I'm an ADMB newbie, using the R2admb package in R to interface with ADMB.

Here's my current dilemma:

In my model, I have several parameters that are initialized and
estimated (e.g., alog, ylog), as well as functions of those parameters
(e.g., d = 1 - exp(-exp(alog + ylog)) ). The initialized parameters
are on the log and logit scale, while the functions of parameters are
on the proportion scale. The results I'm currently getting show
standard errors of some of the initialized parameters (e.g., alog,
ylog) as being absurdly large and equal, but by using "sdreport" for
the functions of those parameters (e.g., d), the estimated standard
errors on the proportion scale are correct and match standard errors
estimated by other maximum likelihood routines (e.g., mle2). I also
calculated variances on the proportion scale by hand via the delta
method and sure enough, when using the standard errors for the
initialized parameters, the variances on the proportion scale were
very wrong (approx 0).

Another related issue is that the variance-covariance matrix is badly
behaved (a lot of covariance estimates of 1 and -1).

The attached zip file contains the R code (data included), the .tpl
file, the _gen.tpl file, as well as other files created by R2admb
(.dat, .pin, .std). Any insight would be much appreciated.

Thanks!
Shanae


-- 
--
Shanae Allen
National Marine Fisheries Service
110 Shaffer Rd.
Santa Cruz, CA 95060
Phone: (831) 420-3970
Email: shanae.allen at noaa.gov
Website: http://swfsc.noaa.gov/SalmonAssessment/
-------------- next part --------------
A non-text attachment was scrubbed...
Name: natmortunbnd.zip
Type: application/zip
Size: 27453 bytes
Desc: not available
URL: <http://lists.admb-project.org/pipermail/users/attachments/20121018/b20e6aa5/attachment.zip>