Probs with running Mark in batch mode

[pablo]

Hello,
I have been running multi-state models using the simulated annealing algorithm that take a rather long-time to converge, and thus I've been wanting to save a list of models to be run in batch mode overnight. While I can create and save models having certain simple features, I have been unable to create models wherein my chosen options of "simmulated annealing algorithm" and parameter-specific link functions are kept. It seems that these choices are somewhat erased when the models are saved, and thus these are run with the default options (logit link and Newton-Raphson).
All help will be appreciated.
Cheers

Pablo

[cooch]

Hello,
I have been running multi-state models using the simulated annealing algorithm that take a rather long-time to converge,

While I'll have to leave the 'bug' you've described for Gary to answer/fix (i.e., the selected optimization technique not being saved...), this is a good opportunity to point out that the simulated annealing method Pablo is talking about is the method used if you select 'Alternate optimization' in the numuerical estimation window (its one of the radio-boxes on the right hand side).

Simulated annealing is a generalization of a Monte Carlo method for examining and solving n-dimensional state equations - and was first described ( I think) by Metropolis et al. (1953) - he of Metropolis-Hasting fame for you Bayesians. Originally described for application to the rate at which crystals annealed - but I think I slept through that part of my physical chemistry class. Its a standard entry in any numerical recipe book you might lay your hands on.

OK, why do you care?

1. good news - simulated annealing is much less likely to suffer the problems of local, non-global minima. Such local minima are surprsingly (annoyingly) common in MS models, for reasons which aren't entirely well-known (see recent work by Giminez).

2. bad news - simulated annealing is MUCH slower than basic Newton-Raphson (or the equivalent). Try it yourself and see. To really test how much of a machine you have on your desk, use SA and profile likelihood-based CI's.

So, you get what you pay for (cost = time). Whenever I get 'weirdness' in models, especially MS models, I typically run SA just to confirm that the 'problem' isn't a local minima.

[gwhite]

Pablo:
I don't know what you are doing, but my version both runs the alternate optimization routine and uses the user-specified links when saved. Are you sure you are saving the model to be run with these options checked?

Note that you can examine the model created, which is just the input file to the numerical estimation program, by using the full output display option. You should see these options specified in the displayed ASCII text file.

Gary