Posterior distribution for finite mixture model?

[f.christiansen]

Hi,
I am fitting a finite mixture model in secr with g0~h2. The pmix parameter tells me that the proportions of the two latent classes are 90% and 10%. If I am interpreting this correctly, this means that the detection function of 90% of my animals has a certain g0 value, whereas the other 10% have a different g0 value. I can obtain both of these from the predict() function. My question is if it’s possible to find out which animals (sighting ID) belongs to each latent class?

All the best,

Fredrik

[murray.efford]

If I am interpreting this correctly, this means that the detection function of 90% of my animals has a certain g0 value, whereas the other 10% have a different g0 value

Only in a manner of speaking - the finite mixture model is used to represent individual heterogeneity in a tractable way in the model, rather than representing a scientific hypothesis about the structure of that heterogeneity i.e. it is a mistake to take the latent classes literally, and the underlying variation may well be continuous.

[Is it] possible to find out which animals (sighting ID) belongs to each latent class?

Not in secr. I do remember Shirley Pledger at one time talking about sorting individuals into classes, and I guess it's possible, but in view of the above I would tread carefully. One can empirically rank individuals by number of detections, but this does not allow for differential spatial sampling (varying number of opportunities per individual) that would be implicit in a model-based approach.
Murray

[f.christiansen]

Thank you for your quick answer Murray,

I am wondering then, if I would like to estimate the encounter probabilities of my marked individuals in different parts of the study area, which g0 estimate would I then use?

Best,

Fredrik

[murray.efford]

Hi Fredrik
I don't have an answer. Yours is a non-standard application of SECR, and you may need a custom model (this is where MCMC & WinBUGS etc. come in handy). For what it's worth, you can compute the average detection probability from a 2-class mixture as the weighted sum of the class values. But I doubt you should be using finite mixtures.
Murray

[f.christiansen]

Thank you Murray,

Yes I was trying to avoid having to use finite mixtures, because of the problem with the biological interpretation. Still for one of the four years which Im analysing, a g0~h2 model provided much more realistic home range centers, so that why I reluctantly tried to keep working with it. It does make some sense also that that particular year had a finite mixture and not the other years, because in that particular year, a few animals were spotted outside the main "core" area, which I believe is a feeding hot spot. These outsiders might have been animals just travelling around, which happened to be spotted by the research boat as it went through these same areas. However such travelling animals were not spotted in other years because of the spatial coverage of the surveys. So while the other years mainly represent animal engaged in feeding activities, some of the animals detected in the particular year might be travelling animals. I dont know if this could be behind the g0~h2 model. And I apologise if Im interpreting too much from it, I just wanted to ask you if you think such differences (yearly variation in spatial sampling effort or detecting animals engaged in different activities) could result in such heterogenity in g0?

Thank you for your answers.

Best,

Fredrik