c-hat & structural data problem

[vin]

I have capture-recapture data for two study sites, small (7 sessions) and big fragments (12 sessions).
At the small fragment site we started to capture birds 5 sessions after the big fragment site, but continued placing mist nets at the same time intervals.
The question of interest is whether or not there is a difference in survival between the fragments.

If I perform a GOF test and calculate c-hat for the fragments using two separate data sets I get sensible values (between 1.3 and 1.7) and non-significant GOF test.
However in order to test the difference I added 5 zero's to each individual capture history of the small fragments data so I could test for a differnce between small and big fragments using one data (.inp) file having two groups.

When I try to estimate c-hat, I get non-sensible values (like -2 using the median c-hat option)
I use a general model Phi (g+t) p (g*t) and restrict the p-values for the first 5 trapping sessions of the small fragment to 0.
I guess the bootstrap and median c-hat procedure does not take this into account?

I would be pleased to hear any suggestions.

thanks in advance,
Vincent

[abreton]

Perhaps you've already found a solution? If not, a couple of thoughts. For the small group you fixed the first five capture probabilities (p) to 0, you should also fix the first four survival probabilities for this group to 0. After fixing these survival parameters to 0, try the GOF test(s) again. Regardless of the outcome, look at the coefficients (betas)for the Phi (g+t) p (g*t) model and make sure that these were successfully estimated by MARK. It may be that some of the betas are not estimable given the data (too sparse). If that was the case, then this would explain why the c-hat statistics are not cooperating, it would also require that you simplify your global model (you'll have to discard Phi (g+t) p (g*t)). Otherwise, if the betas look reasonable and GOF tests for the two group dataset are still not cooperating, then one solution is to apply 1.7 to the two group dataset, you're highest c-hat estimate from assessing GOF separately for each group. When you write up your manuscript, identify that you assessed GOF separately for each group (using MARK I assume?) and report the c-hat estimates (1.3, 1.7). There is no problem with assessing GOF separately for each group, as you've done. You might also consider performing your GOF test in U-CARE, use the two group dataset. Regarding the test of a group effect, my preference, after sorting out the GOF issues, would be to use AICc to "test" for a difference between groups. You'll compare a model with and without groups, but otherwise identical. Calculate an evidence ratio (see Burnham and Anderson 2002 page 77) to quantify support for the group model versus no group model (or vise-versa). If the evidence ratio strongly favors the group model, then I'd assess biological significance of the group effect size.