overdispersion in known-fate models

[ack9f]

Hi,
I am running a known fate survival analysis and am investigating whether or not it is appropriate to estimate overdispersion. I realize the goodness of fit limitations of known fate analysis, but have read several papers where they use a naïve estimate of c-hat for the g*t model to account for overdispersion. Do you think this is appropriate for known-fate data? If so, would this naïve estimate simply be the deviance/df (or the reported c-hat in the specified model output)? I get repeated errors when I attempt to estimate via Bootstrap GOF or c-hat median with saturated and reduced models, but I am assuming that is to be expected considering these methods are not built (and most likely not suitable) for known fate. However, I also have few mortalities so this may be preventing MARK from estimating c-hat with these test options.
Thank you for any advice you can provide.

[bacollier]

Ack9f,
I don't think what you want to do is possible or appropriate for known fate data. See 16.8 in MARK book.

Bret

[ack9f]

I understand that one can not use a traditional goodness of fit test for known fate, but other studies have computed a rough estimate of c-hat when overdispersion was suspected and used it to calculate QAIC. I could try to contact these authors to find out their exact methods for calculating c-hat estimates, but I was also curious to hear if others thought this was a good idea. Does it improve model selection with this naive estimate of c-hat or is it best to leave it alone with possible effects of overdispersion? Thanks in advance for any thoughts on the matter.
-A

[cooch]

I understand that one can not use a traditional goodness of fit test for known fate, but other studies have computed a rough estimate of c-hat when overdispersion was suspected and used it to calculate QAIC. I could try to contact these authors to find out their exact methods for calculating c-hat estimates, but I was also curious to hear if others thought this was a good idea. Does it improve model selection with this naive estimate of c-hat or is it best to leave it alone with possible effects of overdispersion? Thanks in advance for any thoughts on the matter.
-A

Simply because 'others have don'e it' is not always a reliable proxy for 'doing the right thing' (after all, there have been hundreds of studies published that used return rate to 'estimate' survival, the vast majority of which are suspect at best, verging on useless at worst). If you provided some citations, we could have a look, but the whole approach (at least as you've described it) sounds suspicious. At most, what it would do is make the model selection more conservative.

[ack9f]

Thanks for the quick responses. I think my hunger for assessing fit may be causing me to reach beyond what is appropriate. In fact, I had already written my draft with AIC results, when I came across these papers and feared that I had overlooked or misunderstood something. Sounds like the best thing to do is stick with AIC results. Here are two of the papers I was referring to:

Hebblewhite, M., Percy, M. & Serrouya, R. (2003) Black bear (Ursus americanus) survival and demography in the Bow Valley of Banff National Park, Alberta. Biological Conservation, 112, 415-425.

Gusset, M., Ryan, S., Hofmeyr, M., Van Dyk, G., Davies Mostert, H., Graf, J., Owen, C., Szykman, M., Macdonald, D. & Monfort, S. (2008) Efforts going to the dogs? Evaluating attempts to re introduce endangered wild dogs in South Africa. Journal of Applied Ecology, 45, 100-108.

Thank you again.
-A

[bacollier]

-A,
Putting on my smart alec cap: Apparently the answer is G.C. White, personal communication as that is what those authors did. Taking off my cap...

What are you assuming is contributing to the over-dispersion you are wanting to account for? non-independence of fates of radio-tagged individuals?

Bret

[ack9f]

I did radio-tag siblings, which could cause a lack of independence. I completed a chi square test to check for independence within broods based on methods from another study on goshawks (with Gary White personal communication cited ) and concluded siblings acted independently. I just wanted to make sure that I didn't have any other viable options for assessing model fit.