Pushing MARK to the limit

[OlsonD]

I would like to know what are the maximum number of occaisions and strata you can use in the Multi-strata models. I tried running a 48 occasion and 12 strata model to look at seasonal swan movement and MARK didn't like it. Can anyone help?

Thanks in advance
Dave

[tomo_eguchi]

Dave,

With that many occasions and strata, you must have tons of parameters. Do you have enough data for that many parameters? MLE cannot be found (numerically) if the likelihood is too flat over the parameter space. Just my two cents.

Tomo

[Jochen]

Hi,
having the same problem with 115 occasions, but only 6 states and (more importantly) transitions only possible between a subset of the states (meaning that the number of parameters in my model set would be rather limited), I wonder what was the answer to the original question? On a 4 GB RAM W7 machine the most recent MARK build chokes when opening the "Run pre-defined models" window, or copying a single PIM to the clipboard. The latter is a little surprising since Office packages copy and paste even larger spreadsheets. Any information if this is due to built-in limitations of MARK, or could be handled with better hardware, would be appreciated.

[bacollier]

All,
I hit the limit of MARK using Multi-strata models with fewer encounter histories that you have (10 years with 26 strata, for several models when I set them up originally I had >34,000 uniquely estimable parameters, which is insane). Gary advised me offlist to reduce the parameter space using the PIMS by setting all the parameter that are fixed to the same value, then pull them all in as parameter one, followed by trying to just getting a constant model for the other parameters up and running and building off that.

So, probably your first bet is what Tomo addressed, how many parameters are you trying to estimate relative to the data? Then I would check on whether you have set all the constraints appropriately and are not trying to accidentally trying to estimate thousands of parameters. Also, I remember reading it somewhere (Mark or RMark guide/list-Evan/Jeff, help me out here?), but I think that the mlogit() commonly used with multistrata models does not allow for simplifying the number of parameters, but you can change that link to for the estimates to a logit which I think will allow you to reduce the parameter numbers?

Note: I worked on this on my end on a Linux machine with Mark via RMark.

Bret

[jlaake]

Bret is thinking about RMark which uses mlogit by default for multinomial parameters like Psi. I think MARK uses logit by default. The problem is likely to be in the MARK interface but only Gary can comment on that. You may want to consider using RMark as the interface because it simplifies the PIMS prior to running the model but you may still encounter issues unless use the logit for Psi. As Bret said, the parameters for mlogit cannot be simpified due to the way MARK constructs the probabilities for those parameters. It could be changed in MARK but that is up to Gary. I have no idea how much work that would entail.

--jeff