GOF Multi-scale, Multi-method model

[smit8051]

Hi all,

I am trying to evaluate GOF for a multi-scale, multi-method model in Mark. Unfortunately, I have a fairly small sample size for this model (11 sites with 36 sampling events and 2 sampling methods at each event). Based on the simple c-hat estimate provided with MARK’s default model output (deviance/ deviance degrees of freedom) my model is overdispersed (c-hat=11.24). However, I have a small sample size as mentioned earlier and I have read that this method may not be the most appropriate means to evaluate overdispersion with low numbers of samples. I have also modeled these data in Presence and receive the exact same parameter estimates for this model. However, I receive an estimated c-hat value of 1.8048 for this model in program Presence. I believe that Presence uses the MacKenzie and Bailey parametric bootstrap method to estimate c-hat and the probability of the observed test statistic. I have tried to use the GOF bootstrap in Mark to draw a similar bootstrap, but I get the following message, “The bootstrap GOF procedure only works for known fate, live recaptures, multistrata, and dead recoveries at the moment”. I receive a similar error when I try to get a median c-hat estimate. Does anyone have any ideas of how I could go about getting an estimate of GOF for these multi-scale models? I would appreciate any help you might be able to provide me.

Many thanks,
Chris

[abreton]

Have a look at,
http://www.phidot.org/software/mark/doc ... /chap5.pdf

Page 5-6, bottom of the page. You'll find a thorough explanation of the pitfalls of the

deviance/ deviance degrees of freedom

method.

Regarding,

I receive an estimated c-hat value of 1.8048 for this model in program Presence.

As you noted, the two c-hat estimates are different because the two programs are using different methods. Ignore the value you pulled from MARK (see page 5-6) ... and start with the value provided by Program Presence. The estimate you're getting from Presence isn't terminal, I could live with it ... however I'd certainly try to 'model' some of that heterogeneity ... hopefully c-hat would decline to 1-1.5 range.

Regarding the

error

messages from MARK, these aren't errors, instead they're place holders in the development of Program MARK. In future versions of MARK the program will very likely overcome the limitations summarized in the messages that you reported. In the meantime, why not use the c-hat estimate from Presence (Mackenzie & Bailey)? Again, I'd try to reduce a c-hat of 1.8 before doing a final analysis ...

Just some thoughts, hoping they're helpful.