reduced MS model & median c-hat

[arpat]

I am working on a multistate model with 4 states, 5 groups, and 43 occasions. Working with a full time dependent model is neither practically possible, nor my interest. I have enough reason (based on biological knowledge and previous CJS analysis) to assume that survival and recapture rates do not vary significantly over time. My main focus is on the group specific and time-specific variation in transition rates.

When I use U-Care (AS model), c-hat is estimated around 0.25-0.30. A serious underdispersion as my data is too sparse for full time dependent model. However, because I am using a reduced model as my base model, I know I shouldn't use the GOF for the AS model, but use the median c-hat for my actual base model {S(g) p(g) Psi(g*t)}(5-29: Gentle Intro).

However, even running the median c-hat for S(g) p(g) Psi(g*t) model is practically not possible. A single run of the model takes >40 hours. I am thinking about using S(g) p(g) Psi(g) as the base model for median c-hat, and just cross my fingers while estimating/testing time effect on transition rates.

I will be grateful if anyone shares thoughts on this problem..

Cheers,

Arpat

[cooch]

However, even running the median c-hat for S(g) p(g) Psi(g*t) model is practically not possible.

So, just let it run for 40 hours. Whats the big deal? Start it Friday afternoon, and its done by Monday AM. So what if an analysis you have to run only once takes a long time? (I'm currently running an MCMC estimation that is projected to finish in 2 weeks...and this on a quad Opteron running at 2-3 times the spped of your basic desktop).

People have to get over this idea that all results should be generated/completed in minutes. In any other field, taking time to get important results is normal. I have a colleague in physics who is running a job on a massively parallel machine (256 compute nodes) that might be done this year...

So, start the estimation, and leave it alone for a day or two. Surely you have other things you can work on between now and then.

[arpat]

Sorry, there was a misunderstanding on my part. I wasn't worried about the 40 hrs or 2 weeks. I thought median c-hat approach had to run the model with each simulation. So, 120 simulations x 40hrs = 200 days scared me a bit naturally.. I guess it does not need to run the model with each simulation and the test is faster than my initial projection..

Thanks for your input..

[cooch]

Sorry, there was a misunderstanding on my part. I wasn't worried about the 40 hrs or 2 weeks. I thought median c-hat approach had to run the model with each simulation. So, 120 simulations x 40hrs = 200 days scared me a bit naturally.. I guess it does not need to run the model with each simulation and the test is faster than my initial projection..

Thanks for your input..

You do a GOF test *once* for the most general model in your candidate model set. The details of the median c-hat approach are completely spelled out in Chapter 5. I'd suggest reading it carefully, and trying the examples yourself.

[arpat]

That's also what I understood from reading Chapter 5 and doing the excersizes. However, I can clearly see in the command window that iterations and percentage count repeats with each simulation. Yesterday, I started the median c-hat on an even simpler model (which takes only 6-7 hrs to run alone). After 24 hours, it is only 4 out of 120 simulations completed.

I think running the median c-hat for my more complex general model {S(g) p(g) Psi(g*t)} model will still take about 200 days, and I better find another solution..