median c-hat and m-logit link - any new ideas?

[converse]

I noticed a post on this topic to the list a few months ago but there weren't any workable suggestions at that time.
I am trying to do GOF with median c-hat. I am comparing different within-season parameterizations of a 4-state multi-state dataset (using the simple MS, robust design, and open robust design). GOF is a bit of a mess - I've been using Pearson's from MSSURVIVRD but there isn't anything very satisfying for the ORDMS and it would be nice to try the new comprehensive method for all 3 data types.
The problem is that I can't do median c-hat simulations with the m-logit link. I tried to get the global model for the simple MS case to converge with a logit link and simulated annealing. I also tried back-transforming the real estimates from the converged model (using the m-logit) back to a logit link and using these as starting values. Both of these approaches gave me "estimates" but the likelihoods were vastly different than using the m-logit (which I also got with simulated annealing and then used these estimates as starting values in the default optimization routine).
Just wondering if anyone has any other thoughts.

[converse]

Regarding my last post. I was finally able to get the MS model with a logit to converge in MARK using a combination of starting values (I transformed the reals back to the logit scale) and simulated annealing. I then was able to run median c-hat on this, and MARK tells me my answer is about 1.37, versus a little more than 2 based on the Pearsons in MSSURVIVRD!
Of course, I then realized that MARK won't (yet??) run the m c-h simulations on RD data, so I'll have to take an alternate approach there.
Looking forward to seeing more developments on this front, Gary.

[cooch]

Regarding my last post. I was finally able to get the MS model with a logit to converge in MARK using a combination of starting values (I transformed the reals back to the logit scale) and simulated annealing. I then was able to run median c-hat on this, and MARK tells me my answer is about 1.37, versus a little more than 2 based on the Pearsons in MSSURVIVRD!
Of course, I then realized that MARK won't (yet??) run the m c-h simulations on RD data, so I'll have to take an alternate approach there.
Looking forward to seeing more developments on this front, Gary.

Glad you got it working. In fact, I was just about to post a reply - pointing out that I'm quite often using precisely the same sequence you describe - I now routinely fit models using simulated annealing - whenever you have lots of parameters, its amazing how many times its able to 'get there from here' when the default optimization routines crash and burn.

On your wish list for things for Gary to work on in his copious spare time is the ability to have 'saved models' be specified with simulated annealing. At present, MARK runs a series of saved models using the default optimization. One of the obvious downsides for SA is how long it takes - which is often strong motivation for building the set of models, and letting them run overnight.

[gwhite]

Evan:
If you save a model with the Alt Opt. Method box checked, the model is then run later with the simulated annealing method, not the usual optimization method. You can verify that by the fact that the SIMANNEAL option is set in the proc estimate statement.

Just be sure to check the box for Alt. Opt. Method.


Gary

[cooch]

Evan:
If you save a model with the Alt Opt. Method box checked, the model is then run later with the simulated annealing method, not the usual optimization method. You can verify that by the fact that the SIMANNEAL option is set in the proc estimate statement.

Just be sure to check the box for Alt. Opt. Method.


Gary

So I see - it *didn't* work on an older version of MARK I was using on another machine.

Thanks...