Inconsistent confidence intervals csolnp / npsol - plots

Posted on
No user picture. CharlesD Joined: 04/30/2013

Should I expect confidence intervals to be working for csolnp at the moment?
I ran a number of fits, generating random start values, using the same data. The following plots of four of the model parameters and confidence intervals show quite some inconsistency / strangeness, I think. Although the likelihood varied from fit to fit, the differences were very slight. Estimates were all from good fits (no code reds / blues).

The x axis of the plot is both parameter (1-4) and within each parameter section is iteration 1-20. scaled likelihood values 1-20 are in the top section of plot.

https://www.dropbox.com/s/elk789dmy43ctmw/csolnp.png?dl=0
https://www.dropbox.com/s/xh8j1wywevswtyj/npsol.png?dl=0

these data files contain lists of the fit objects from the 20 iterations, in case they are of interest.
https://www.dropbox.com/s/jpqn4hsc85z9q7c/csolnp_nobounds_fits.RData?dl=0
https://www.dropbox.com/s/1tpdl0keeg1ttzg/npsol_nobounds_fits.RData?dl=0

Replied on Mon, 09/08/2014 - 10:33
Picture of user. RobK Joined: 04/19/2011

Based on what you're reporting here, no, it doesn't look like confidence intervals are working properly with CSOLNP.

EDIT: Have you tried the new (third) beta (or built from the head of the repository recently)?