Dependence on both previous and current states

[Aline]

Is it possible in RMark to have encounter probabilities dependent on both previous state and current state? Any thoughts on how I might go about it?

I have run Gof tests for a multistate model in UCare, and they seem to be based on the JMV model, so I think I need to run this model for comparison with my (presumably) more appropriate models.

Thanks

[jlaake]

By encounter probabilities, I assume you want to make p to depend on current state and previous state. I don't think that is possible in the multistate model MARK. If it is possible,someone correct me and I'll work out how to setup the model for MARK with RMark. If you are talking about state transitions, that is quite simple.

[cooch]

By encounter probabilities, I assume you want to make p to depend on current state and previous state. I don't think that is possible in the multistate model MARK. If it is possible,someone correct me and I'll work out how to setup the model for MARK with RMark. If you are talking about state transitions, that is quite simple.

No, you're correct. A model where encounter depends on states at start and end of the interval was describd by Brownie et al. (1993), and is known as the JMV model (also discussed in Williams, Nichols & Conroy). It didn't ee much use, but was resurrected by Pradel et al. as a fundamental componet of the MS goodness of fit testing they implement in U-CARE.

From Chapter 10:

What is this JMV model, and how does it relate to the AS (Arnason-Schwarz) model? The JMV model (Brownie et al. 1993) differs from the typical AS model in that it permits the capture probability for time (i+1) to depend on the state at periods (i) and (i+1) (whereas the AS model permits the encounter probability to depend only on current state, and time). Thus, the AS model is in fact a special (reduced) case of the more general JMV model. As with program RELEASE, a fully efficient GOF test for the JMV model is based on the property that all animals present at any given time on the same site behave the same...

[jlaake]

Good to know I haven't completely lost it "yet". This type of model could be constructed as a second-order hidden Markov model. Such a model would also be useful for state transitions at time t+1 that depended on state at time t-1 and t, such as birds and whales that use summer and wintering areas and captures occur at both areas. It isn't that difficult to do but I just haven't had time to implement it. It creates lots of possible parameters so you have to be careful. I wonder whether it is possible in ESurge.

I find it odd that the capture probability in one state would influence the capture probability in the next state. Is it possible that this is simply individual heterogeneity in p that is not being modeled? You can make capture at time t a function of 0/1 at time t-1 (trap dependence) but not a function of its prior state. Not sure if that makes sense for the MS model though.

regards --jeff

[cooch]

Good to know I haven't completely lost it "yet". This type of model could be constructed as a second-order hidden Markov model. Such a model would also be useful for state transitions at time t+1 that depended on state at time t-1 and t, such as birds and whales that use summer and wintering areas and captures occur at both areas. It isn't that difficult to do but I just haven't had time to implement it. It creates lots of possible parameters so you have to be careful. I wonder whether it is possible in ESurge.

I'm fairly certain it is. Comes up a lot in disease modeling. Paul Conn and I used E-SURGE for the work we did.

I find it odd that the capture probability in one state would influence the capture probability in the next state. Is it possible that this is simply individual heterogeneity in p that is not being modeled? You can make capture at time t a function of 0/1 at time t-1 (trap dependence) but not a function of its prior state. Not sure if that makes sense for the MS model though.

regards --jeff

[Aline]

Thank you both for your answers. I'm new to this and might be making things over complicated, but here's my logic. Please correct me if I've misunderstood how things work.

Running a goodness of fit test establishes whether the general model (in U-Care, that would be JMV) fits the data. Other models that are tested against this model and found to explain the data better can then be assumed also to fit the data.

Doesn't that then require that I include the same general model in my model testing? I don't see any reason why detection probability in my system should be dependent on previous state, so I suppose I could argue that that attribute of the model is unlikely to change the fit, but would that really be acceptable? Especially if I have to correct for overdispersion.

I see that U-Care is recommended for Gof testing of multistate models when using Mark, so I'm wondering if I'm overthinking this.