Model fails with different specification of trap-dep states

[jg661]

Hi all,

I'd be really grateful for some advice on a piece of analysis I've been working on.

I have been working on quite a complex multistate model for seabird data with the following set up:
7 events (three locations, two breeding states for each location)
13 states (same as states but with a trap-aware and trap-unaware version of each equivalent event)
3 groups
7 occasions
0 individuals covariates

As my data has trap-dependence, I am using the Pradel & Sanz-Aguilar approach to deal with this. I have three transition matrices: survival, dispersal and detection (trap-aware/unaware process). My event probability matrix is not active as it is synonymous with the detection matrix, so I have used i as the GEMACO syntax to set this.

I have previously run versions of this model, only obtaining estimates of survival and dispersal for the trap aware birds, by using the GEMACO syntax to specify only the trap-aware states. These models have run fine and seem to converge and give sensible results. However, I have realised that it is preferable to generate survival and dispersal estimates for all birds, not just those in the trap-aware states, so have rerun exactly the same models, but these are failing. The model runs fine with the amended GEMACO syntax for survival, but fails when I try and amend the dispersal syntax. For example:

where:
1 - non breeder at site one (trap-aware)
2 - non-breeder at site one (trap-unaware)
3 - non breeder at site two (trap-aware)
4 - non-breeder at site two (trap-unaware)
5 - non breeder at site three (trap-aware)
6 - non-breeder at site three (trap-unaware)
7 - breeder at site one (trap-aware)
8 - breeder at site one (trap-unaware)
9 - breeder at site two (trap-aware)
10 - breeder at site two (trap-unaware)
11 - breeder at site three (trap-aware)
12 - breeder at site three (trap-unaware)

and where:

the first set of [] specifies the probability to remain a non-breeder regardless of location, the second set of [] specifies the probability of a non-breeder becoming a non-breeder at any of the other locations; and the third, fourth and fifth set of [] specify the probability of a current breeder becoming a breeder at either of the other two locations. Age classes are used so that the probability of a non-breeder younger than a set age becoming a breeder can be set to zero using IMVF.

GEMACO syntax for dispersal of only trap-birds that runs:

[f(1 3 5).to(1 3 5)]+[f(1, 3, 5).[young+ad2].to(7, 9, 11)]+[f(7).[young+ad2].to(7, 9 11)[f(9 10).[young+ad2].to(9 10, 7 8 11 12)]+[f(11 12).[young+ad2].t(11 12, 7 8 9 10)]

GEMACO syntax for dispersal of both trap-aware and -unaware birds that fails:

[f(1 2 3 4 5 6).to(1 2 3 4 5 6)]+[f(1 2, 3 4, 5 6).[young+ad2].to(7 8, 9 10, 11 12)]+
[f(7 8).[young+ad2].to(7 8, 9 10 11 12)+[f(9 10).[young+ad2].to(9 10, 7 8 11 12)]+
[f(11 12).[young+ad2].t(11 12, 7 8 9 10)]

When the second version of the model runs it fails very quickly, I receive warnings that the first 10 histories are incompatible with the model with each run (model set to run 5 times with random initial values), as well as the following errors at the end of the overall run: "Error using matlab.graphics.axis.Axes/set Value must be a vector of type single or double whose values increase" and errors in "barV6", "bar", "hist", "menumuse>Run_Callback", "gui_mainfcn" and "matlab.graphics.internal.figfile.FigFile".

I'd be grateful for any insight or advice regarding these issues, and am happy to provide more information if needed.

Many thanks in advance,

Jennifer

[simone77]

Hello Jennifer,

I'm not sure how much assistance I can provide here, but I have a few considerations. As you've noted, this is indeed a complex multistate model. From my professional experience, such complexity typically demands a dataset of considerable size, potentially surpassing the scale encountered in typical capture-recapture modeling. Consequently, there's the possibility that your dataset might not furnish sufficient information to fully support the demands of your modelling approach.

With that in mind, it would help undestand your case a more detailed elucidation of the structure employed in GEPAT. For example, You have not given details on GEPAT and say almost nothing about Initial State and Event stuctures. While I infer that you may be aligning your approach with the model presented in Pradel & Sanz-Aguilar's 2012 paper, a more comprehensive exposition of its implementation would be beneficial in furthering our comprehension of your analysis.

I do not understand the absence of a state-transition GEPAT for the breeding component. I would expect the presence of matrices accounting for survival, breeding, and dispersal processes. Moreover, the categorization of three distinct groups remains somewhat enigmatic; Do these groups correspond to age-at-first-capture dynamics (and why ad1 and ad2)? How you're addressing the deterministic age transitions of individuals initially captured as juveniles, subsequently becoming adults? If you do not know much of this, it might be helpful to explore the concept of "time-since-marking" within the MARK book (it is completely helpful also for a model built in E-SURGE).

Lastly, I would underscore that successful execution of a model doesn't inherently confer accuracy or meaningfulness. In practice, it's not uncommon for a model or a code snippet (e.g. in R or Python) to execute seamlessly, producing aesthetically appealing results, yet ultimately yield conclusions that are not reflective of the underlying data or biological reality.

Simone