Using U-CARE with missing site visits

[josephine]

Hello!
I'm using a multistate model with live recaptures in program MARK to model the detection and transition of roost trees from standing to fallen (survival is fixed at 1). I have data on roost trees (~260) that were discovered annually over a 12-year period and surveys of these trees that were conducted annually for 3 years at the end of this period. Limited time meant all trees couldn't be revisited during each survey, so all trees were surveyed at least once with some surveyed 2 or 3 times. The data was not collected with a MARK analysis in mind, so the encounter histories are rather strange with unequal time intervals between when trees were discovered and the surveys and many missing site visits. Below is an example of my encounter histories:

Encounter Histories
.1.........10
....1......1.
....1......0.
...1......11.
1..........1.
.........12..

Example of meaning
Tree was discovered in 2013, surveyed in 2023 and found standing, and not located during the survey in 2024

Program U-CARE seems to be the best way to evaluate the goodness of fit for multistate models, but it doesn't allow special characters. Does anyone know a workaround for this given how odd my encounter histories are? Any advice on how to proceed is greatly appreciated. Thanks!

[cooch]

Sorry to be the bearer of bad news, but you're out of luck. Not only do you have more 'dots' than 'data', but unequal intervals (which is what this situations amounts to) creates major problems for MS models generally (specifically, makes parameter difficult to interpret without strong assumptions).

And, to your initial question -- no, I don't know of any workaround for U-CARE (or GOF testing in general) for the sort of data you seem to have (or, rather, lack of data).

[josephine]

Thanks for the response. I've actually created a model structure in MARK already which seems to be performing relatively well, or the parameter estimates are at least intuitive and seem reasonable.

I did make some pretty strong assumptions:
1. Survival fixed at 1 - trees were not leaving the landscape
2. Detection only possible during survey years, fixed at 0 when surveys were not being conducted
3. Detection predicted by a single parameter but allowed to vary based on tree state (standing or fallen) and a grouping variable (live or dead)
4. Only predicting the probability of a dead tree transitioning from standing to fallen, others fixed at 0 because they were not observed

This is quite simplified compared to how many parameters could be modeled, but it's appropriate given the circumstances and the assumptions are well-founded. The prospects for GOF testing in U-CARE indeed seem bleak, but I welcome any alternative suggestions. I imagine this isn't the first time (or last) that limited data with missing site visits is being analyzed in MARK.