transience and trap dependence

[Mark Trinder]

Hi all

we've been using Ucare to try and find the appropriate starting model for a multistrata analysis of a guillemot dataset. We are working with 3 states: chick, non-breeder and breeder, with all birds ringed as chicks. The GOF tests indicate a high level of transience since a large proportion of the ringed chicks are never seen again, and trap dependence for all 3 states, caused by a range of issues: chicks which do return (as either non-breeders or breeders) do so in their third year at the earliest, and many not until their fourth year (eg 100223 is a typical EH); most re-sighting effort has taken place at a breeding colony, so not surprisingly breeding adults (particularly round the edges of the colony) have good re-sighting histories, which also skews the encounter histories, etc etc.

We've been battling our way through the available sources of advice for the multistate GOF, which is pretty hard going , however a best guess approach would seem to be;

a) split for trap dependence and
b) use the dummy age parameterising trick for transience

The main question is whether it is appropriate to apply not only one, but both of these two types of correction to a multi-state analysis (in fact it would be nice to know if a combined approach such as this is appropriate for single state models too).

A further complication is the question of whether we should correct survival or state transitions for the dummy age manipulation (for the transient problem)......

Assuming any or all of the above approaches are valid, it's also not obvious which combination of tests to use for calculating c-hat for multistate modelling if you have undertaken the sorts of manipulations described above.

(basically this is a plea for Evan to write a comprehenive guide to GOF for multistate models to complement all the other excellent help he has produced! - only can you do it by tomorrow please

cheers

Mark
(& Steve Votier)

[cooch]

<snip>...

A further complication is the question of whether we should correct survival or state transitions for the dummy age manipulation (for the transient problem)......

Assuming any or all of the above approaches are valid, it's also not obvious which combination of tests to use for calculating c-hat for multistate modelling if you have undertaken the sorts of manipulations described above.

(basically this is a plea for Evan to write a comprehenive guide to GOF for multistate models to complement all the other excellent help he has produced! - only can you do it by tomorrow please

If the issue was transience only, then it seems pretty clear from some work we've done at this end (paper still under review) that the presence of transients does not bias estimates of transition rates - which should make some sense if you're separately estimating survival and movement. Transience is known to negatively bias estimates of survival, but if survival is temporaly separated from movement, then this isn't a particular problem, as far as we can determine.

As for combining trap effects, and transients, and GOF for the two together - not sure. Basically, GOF for MS models in general is still somewhat of an open question - U-CARE is a major step forward, but presumes the general model is fully time-dependent, and essentially a MS analogue of CJS.

The other approach is to use the median c-hat. I've found that for 'well-behaved' CJS-MS models, the median c-hat gives basically the same results as does U-CARE. Given that, it seems reasonable to look at the median c-hat as a way forward, with the caveat that it too is a work in progress.

[cooch]

If using U-CARE, then you can use one of the component GOF tests (TEST3G.SR) to test for evidence of lack of fit to the fully time-dependent model due to the presence of TSM effects (R. Pradel, pers. comm.). Basically, this is analogous to the similar trick of partitioning test 2 and test 3 for TSM (age) effects in standard CJS models.

The primary citation which should be consulted here is

Pradel,R., Wintrebert,C.M.A. & Gimenez,O. (2003) A proposal for a goodness-of-fit test to the Arnason-Schwarz multisite capture-recapture model. Biometrics, 59, 43-53

Not a 'light read', but for the moment, the essential starting point.

[Mark Trinder]

cheers for that rapid reply Evan.

One quick follow up to your suggestion of using the median c-hat - we have a pretty strong hunch that trap dependence is a serious issue, so were planning to start from a data file split for trap dependence, but when you run the median c-hat routine it throws up a warning message that any encounter histories with removals on re-capture (ie the majority of the trap dep EH file) will be effectively ignored...

so is the appropriate approach to use a normal inp file and let the median c-hat 'deal' with the trap dependency issues? or should we estimate c-hat from the unmodified dataset and then use that with the trap-dep data?

cheers
mark and steve

[cooch]

One quick follow up to your suggestion of using the median c-hat - we have a pretty strong hunch that trap dependence is a serious issue, so were planning to start from a data file split for trap dependence, but when you run the median c-hat routine it throws up a warning message that any encounter histories with removals on re-capture (ie the majority of the trap dep EH file) will be effectively ignored...

so is the appropriate approach to use a normal inp file and let the median c-hat 'deal' with the trap dependency issues? or should we estimate c-hat from the unmodified dataset and then use that with the trap-dep data?

I suspect this is a question I'll have to let Gary field, since he has deeper insights into how the median c-hat is being applied (obviously).