Multi-state model with memory

Forum for discussion of general questions related to study design and/or analysis of existing data - software neutral.

Re: Multi-state model with memory

Postby jlaake » Tue Mar 22, 2022 8:00 pm

With regard to ESurge which I don't know much about, as long as it handles hmm first order generically then the way second order is handle is to make pairs of strata combinations. So you'll go from an 8 by 8 transition matrix to a 30 by 30 matrix (5*3 +3*5) so handling that manually will be excruciating. I haven't given it much thought but MARK has a generic HMM model and it is supported by RMark. The design data and structure will not be automatic and not sure how or whether it will work. Worth looking into though.
jlaake
 
Posts: 1417
Joined: Fri May 12, 2006 12:50 pm
Location: Escondido, CA

Re: Multi-state model with memory

Postby B.K. Sandercock » Wed Mar 23, 2022 4:04 am

Here's an example of a memory model fit with SURVIV for a large dataset of Canada geese.

Hestbeck, J.B., Nichols, J.D. and Malecki, R.A. (1991), Estimates of Movement and Site Fidelity Using Mark-Resight Data of Wintering Canada Geese. Ecology, 72: 523-533.
PDF available at JSTOR: https://www.jstor.org/stable/pdf/2937193.pdf
B.K. Sandercock
 
Posts: 48
Joined: Mon Jun 02, 2003 4:18 pm
Location: Norwegian Institute of Nature Research

Re: Multi-state model with memory

Postby awan » Wed Mar 23, 2022 4:32 am

Hi Jeff, regarding the data, the oystercatcher data actually goes much further back, but with seasonal time steps, I already found that 20 years (40 time intervals) was creating memory issues. However the study only truly expanded to all geographic states from 2008 onwards, hence why I limit the data to the last 10 years (20 time intervals). I had simplified the PIM for psi to "time" for the adult-only models, but the years that include juveniles include different levels of grouping where I couldn't simplify the PIM and for these models one of the limitations is run-time. And I have been using RMark but I didn't realise that the number of individuals isn't a limitation but from what you write it is probably something I should've known. Thanks for following up
awan
 
Posts: 16
Joined: Fri Feb 10, 2017 10:52 am

Re: Multi-state model with memory

Postby awan » Wed Mar 23, 2022 5:31 am

B.K. Sandercock wrote:Here's an example of a memory model fit with SURVIV for a large dataset of Canada geese.

Hestbeck, J.B., Nichols, J.D. and Malecki, R.A. (1991), Estimates of Movement and Site Fidelity Using Mark-Resight Data of Wintering Canada Geese. Ecology, 72: 523-533.
PDF available at JSTOR: https://www.jstor.org/stable/pdf/2937193.pdf


Hi Brett, thanks for the article. That is indeed the type of model I am after.

Armed with new search terms, I thought I'd share another I article I found providing a comparison of memory models (that also uses the goose data), and with supplements describing how the models were run in E-Surge. https://doi.org/10.1002/ece3.1037. It seems like it time to start learning a new software and to turn towards E-Surge as suggested by others. Once again thank you for the helpful replies all.
awan
 
Posts: 16
Joined: Fri Feb 10, 2017 10:52 am

Re: Multi-state model with memory

Postby egc » Wed Mar 23, 2022 7:32 am

B.K. Sandercock wrote:Here's an example of a memory model fit with SURVIV for a large dataset of Canada geese.
.

Key point being large, with very high encounter probability. It will not surprise you that a fair number of the applications of 'memory' models are applied to studies of goose populations: high density nesting, extremely high female philopatry, easy to catch, both live encounter and dead recovery data, virtually none of that pesky 'spatially explicit' stuff to worry about (geese don't have territories, and perhaps more than most species, nicely match the statistical starting assumption that your organism is a randomly moving Brownian particle). The goose data I worked with for the bulk of my career was about 75-100K marked individuals, with encounter probabilities (at least in the early years) of p > 0.5. [Of course, this pales compared to Emmanuelle Cam's kittiwake data -- when you have encounter probabilities that approach 1.0, as her study does, combined with relatively large sample sizes, you can fit whatever models you want.]

There is a long-standing tradition of the 'smart folks' coming up with clever and interesting models, which are 'demonstrated' in some paper with (typically) 'very good -> perfect data'. Masses read the papers, try to apply the clever idea to their own data, only to find with some frustration that it might not work, because their data aren't equivalent to the frequently near-optimal 'empirical example' data used in the paper describing the models in the first place.

Summarizing -- memory models are 'data hungry', which is absolutely one of the reasons they don't get used much. And, building some of the models can be challenging, since the 'chain of conditional probabilities' (e.g., state A to B this year, conditional on A to A last year, B to A last year, C to A last year...) you might need to model can get large, and complicated, very quickly.
egc
Site Admin
 
Posts: 201
Joined: Thu May 15, 2003 3:25 pm

Previous

Return to analysis & design questions

Who is online

Users browsing this forum: No registered users and 19 guests