GoF for multi-state CR.

[Jake]

Hello,


I have constructed a typical multi-state encounter histories MARK file for individuals in a CR experiment with 15 capture occasions. The MARK file also contains the sex specification of each individual (i.e. two groups male/female). Eg.

011011000000000 1 0;
000210220000000 0 1;
.
.
.

etc.

I tested the GoF for my data set using the SUM of MS test in U-CARE. For both the male and female groups, there appeared to be NO lack of fit with p-values of 0.911 and 0.798 respectively for the AS-model. However, when I pool the groups together and run the SUM of MS test there appears to be a lack of fit with a p-value of 0.000 for the AS-model. The JMV model also showed a lack of fit when groups were pooled.

1. Does this indicate that a gender covariate must be considered in my most general model (i.e. the AS model with a sex covariate) ?

2. Can I still use this data set for MS modeling (with sex as a covariate) even though there appears to be a lack of fit?

3. Should I just look at each group separately and perform MS modeling on each group alone? Then analyze and compare survival, recapture and transition probabilities for each sex?


Thanks.

[cooch]

Hello,


1. Does this indicate that a gender covariate must be considered in my most general model (i.e. the AS model with a sex covariate) ?

Probably - or that the overall heterogeneity tghat exists only shows up when you have a sufficiently large sample size (as you might get by pooling over sexes).

2. Can I still use this data set for MS modeling (with sex as a covariate) even though there appears to be a lack of fit?

Depends on the likely sources of the lack of fit. If the reasons are strcutural, you search for better general model. If they're extra-nomial (i.e., lack of independence between males and females) you can proceed by estimating c-hat and adjusting. This is covered in Chapter 5.

3. Should I just look at each group separately and perform MS modeling on each group alone? Then analyze and compare survival, recapture and transition probabilities for each sex?

Depends on the larger questions.

[cooch]

worth estimating c-hat in a couple of ways - U-CARE, and median c-hat in MARK. Regardless, beware local minima - see MS chapter in MARK book for general discussion of this issue.

[Jake]

Thank you for your comments and suggestions.

[CHOQUET]

Hello,

To my opinion, there is mainly two issues:

either
1) there is too few individuals in each group too detect lack of fit
or
2) either a group effect(i.e. sex) should be consider. Your general model
is in this case phi(site*t*g)p(site*t*g).

You can easily recognize the 1) case in U-CARE if several tests are empty(0,NaN).

Rémi