Correcting for overdispersion in a multistate model

questions concerning analysis/theory using programs M-SURGE, E-SURGE and U-CARE

Correcting for overdispersion in a multistate model

Postby bootzies » Mon Aug 11, 2014 4:01 am

Hello all.

I am hoping to fit a multistate model to my capture-mark-recapture (alive only) dataset, which contains 5800 individual encounter histories over a 30-year period. There are 6 states (representing parasite load) and one grouping variable (age group: adult or juvenile).

I have run the goodness of fit tests for the JMV model in U-CARE and these give significant results for the adult group (the adults make up the vast majority of the individuals captured). I am unsure of how to proceed now.

In table 13 of the U-CARE manual section on multistate GoF tests (p.34-35) it states that if you have significant results in your GoF tests, you should ‘start from corresponding model with over-dispersion coefficient’. I am not 100 % sure what this statement means, but I assume it means that I should try and fit the general model (i.e. JMV), but somehow adjust the c-hat value. And this is where I am stuck…… I am unsure how to estimate an appropriate c-hat value for multistate models (I don’t think it’s possible to do this using the median-c bootstrap approach or other approaches offered for single-state models). Is there a way to use the U-CARE software to estimate an appropriate over-dispersion correction factor for the dataset?

Any help would be much appreciated!

Thanks,

Alice
bootzies
 
Posts: 21
Joined: Wed Jan 29, 2014 2:31 am

Re: Correcting for overdispersion in a multistate model

Postby CHOQUET » Mon Aug 18, 2014 9:26 am

How to handle overdispersion is not easy!!
The first think is to identify the main effect for overdispersion and build a model
with such a effect. See the various paper on that subject for capture-recapture.
Each year, we organise a workshop where GOF is one of the main subject.
CHOQUET
 
Posts: 211
Joined: Thu Nov 24, 2005 4:58 am
Location: CEFE, Montpellier, FRANCE.

Re: Correcting for overdispersion in a multistate model

Postby Guillaume Souchay » Mon Oct 27, 2014 11:03 am

Hi Alice,

Do you still have trouble with the goodness-of-fit tests for your analysis and the use of an overdispersion coefficient?

Cheers,
Guillaume
Guillaume Souchay
 
Posts: 64
Joined: Wed Feb 27, 2008 5:20 am
Location: Nantes, FRANCE

Re: Correcting for overdispersion in a multistate model

Postby bootzies » Mon Oct 27, 2014 7:57 pm

Hi Guillaume,

Yes, I do to some extent!

I did manage to run the median-C simulation test in MARK on my multistate model - but this gave strange results. It gave a c-hat of ~1.2, which is only indicative of slight overdispersion, however the U-CARE tests for the JMV model indicated serious deviation from the expected structure (very significant tests with p-values <0.001). I am still very confused as to why/how the two tests can give such different results. I know that they don't test in the same way, however I would expect that they ought to generally give similar results!

I am now fitting a multistate model with a transient effect. It looks as though there are transient individuals in the population, which could be the reason for lack of fit (and needs to be dealt with, as it could be biasing the survival estimates).

Any advice would be gratefully received!

Cheers,

Alice
bootzies
 
Posts: 21
Joined: Wed Jan 29, 2014 2:31 am

Re: Correcting for overdispersion in a multistate model

Postby Guillaume Souchay » Tue Oct 28, 2014 3:50 am

Hi Alice,

If you want to understand why you have some difference in the result, it can be good to understand what does each program. I used more often U-CARE, thus I can try to explain what it does.
The program U-CARE doesn't compute a c-hat for you, it only provides tests (either for CJS-based or for JMV-based models). Based on the p-values and the degree of freedom, then, you can compute a c-hat. Perhaps, the p-value providing by U-CARE in your case refers to the transient effect, but the associated test gives only a low c-hat value ...
Did you read either the Chapter 5 of the Mark book or the U-CARE manual? If you want to understand what happen in your data, you can just perform each test separately (Tests WBWA, 3G and Tests M) to see if you have memory, transience and trap-dependence respectively.
Another concern is that the JVM model refers to multi-site model. In your case, you have a multi-state model to describe biological states of the individual (linked to the parasite load). In that case, perhaps, distinguishing between states for the goodness-of-fit test is not necessary. Maybe, all the transition are not possible in your model, or there are some constraints that could induce some tests to be significant.
If you are more concerned by the fact that individuals are reencountered again or not, and if the recapture is supposed to be the same for each biological state, you could recode your data (using U-CARE option) to be then in a single-state case (detected or not) and run a CJS GoF-test.

Cheers,

Guillaume
Guillaume Souchay
 
Posts: 64
Joined: Wed Feb 27, 2008 5:20 am
Location: Nantes, FRANCE


Return to analysis help

Who is online

Users browsing this forum: No registered users and 9 guests