How to estimate c-hat for an age-structured model

[bootzies]

Hello,

I am interested in survival of seabirds as a function of temporal covariates (e.g. weather and availability of fish stocks) and age class (using 3 age groupings: immature, adult, old) - with individuals transitioning through the age classes over time (i.e. age-structured model).

I am confused as to how to best to estimate c-hat and/or measure model fit for an age-structured model. As I understand it, the RELEASE and U-CARE programs can deal with grouping variables - so I could account for age-dependent survival in these tests based only on the age at first capture (which is always 0, as all birds are banded as chicks). But, I want to test GOF for a structured model where survival probability is dependent on age class with individuals aging through time. I can't see any way of doing this in RELEASE and U-CARE tests.... Is there?

When I run the standard (un-grouped) GOF tests in U-CARE/RELEASE the results indicate high over-dispersion. I think that this is (at least in part) because these tests are not accounting for the influence of age-structure on survival. Younger birds are much more susceptible to weather, predation and food shortages, and the oldest birds die off due to senescence (so I would expect the middle class to have the highest survival).

How can I test my data for overdispersion/GOF to an age-structured model?

I am using RMark to code the model and edit the design matrices for age transitions. I have tried to import the resulting model into MARK so that I can run bootstrap or median-c tests. I used the appropriate code to generate an Rinp and then the 'import RMark' option in MARK. However, I have never managed to achieve success with this import procedure! I either get an error message, or MARK hangs and I have to force quit (after waiting hours for something to happen).

Any help or advice would be much appreciated.

Thanks,

Alice

[cooch]

I am confused as to how to best to estimate c-hat and/or measure model fit for an age-structured model. As I understand it, the RELEASE and U-CARE programs can deal with grouping variables - so I could account for age-dependent survival in these tests based only on the age at first capture (which is always 0, as all birds are banded as chicks). But, I want to test GOF for a structured model where survival probability is dependent on age class with individuals aging through time. I can't see any way of doing this in RELEASE and U-CARE tests.... Is there?

Not easily -- for 2-age class models, yes, but not >2.

When I run the standard (un-grouped) GOF tests in U-CARE/RELEASE the results indicate high over-dispersion. I think that this is (at least in part) because these tests are not accounting for the influence of age-structure on survival.

This is undoubtedly correct.

How can I test my data for overdispersion/GOF to an age-structured model?

If all you want is an estimate of c-hat (as a parameter), without the fine-grained assessment as to 'where/why' you have lack of fit, one option is to simply look at the Fletcher c-hat statistic. See section 5.8 in Chapter 5.

I am using RMark to code the model and edit the design matrices for age transitions. I have tried to import the resulting model into MARK so that I can run bootstrap or median-c tests. I used the appropriate code to generate an Rinp and then the 'import RMark' option in MARK. However, I have never managed to achieve success with this import procedure! I either get an error message, or MARK hangs and I have to force quit (after waiting hours for something to happen).

Jeff will need to field that. 9/10 times, this problems seems to be related to either (i) permissions, (ii) the version of either RMark or MARK being out-of-synch, or both.

[bootzies]

Thanks very much for your response. I will take a look at the Fletcher c-hat statistic.
Any information from Jeff on what may be causing the issues with importing models from RMark would be gratefully received...
Alice

[jlaake]

Sorry I missed that at the bottom of the message. I had thought that Evan had addressed the question and stopped reading. Probably best to avoid a single posting that is composed of 2 different thoughts. A posting that was titled "problem importing into MARK from RMark" would have gotten my immediate attention. The re-posting was the only reason I caught this.

I did just fix a problem with export.MARK but it was for models in which the encounter history had 2 characters like the Both/Barker models etc. It was outputting an incorrect number of occasions. What model are you using? If you look at the code in export.MARK it is really quite simple. It creates the .Rinp and .inp files (both are text) and then renames the .out,.vcv,.res files to have a .tmp extension (function export.model) such that they will be recognized by MARK for importing. Look at the .Rinp and see if the number of occasions (3rd line) is correct. Also, delete all of the .tmp files in the directory and then try importing. That should create the .dbf with an empty .fpt file (no results). If it hangs then, then there is an incompatibility between RMark and MARK and Gary and I will have to work it out. If there are lots of .tmp files then it may just be taking a long time loading them. Try specifying a single global model to export and see if it can be imported.

Without knowing more or seeing the .Rinp file it is hard for me to diagnose any further.

--jeff