Hi,
I know it is already summer time and these are not good times to post new topics but unfortunately I have this trouble right now...hope to be lucky...
I am working with resightings of a species of bird and I am interested to estimate transience and survival. In my case study one individual might be randomly: (i) molecularly sexed (no uncertainty), (ii) visually sexed (uncertainty), (iii) molecularly and visually sexed, (iv) neither molecularly nor visually sexed.
Once it was visually sexed, it might be sexed from one team (Monitoring Team nº1) or another (Monitoring Team nº2).
I have modeled my data in a very similar fashion to that Genovart, Pradel and Oro (2012) used in Exploiting uncertain ecological fieldwork data with multi-event capture–recapture modelling: an example with bird sex assignment. They partitioned the transition in two steps (transience and survival) and encounter in several steps to take into account the sex uncertainty.
Here it is a one-page PDF with the states, events, groups, shortcuts, and GEPAT structure I have used for these analyses.
All the models in my candidate set were run with Multiple Random option =5 (to minimize local minima risk), and tolerance to parameter change and tolerance on gradient= 1e-009 (to ensure convergence).
Results indicate no trouble with parameter identifiability.
There are two things that look strange (to me at least):
1. E-SURGE results don't agree at all with those I get in U-CARE.
In U-CARE I found a strong transient effect in males and no transient effect in females and in E-SURGE almost all the models find just the contrary (but see below). I have checked (and tested) data arrangements for both analyses (E-SURGE and U-CARE) one million times and I am sure there are no problems there.
2. In E-SURGE I have found that a different parameterization of Survival had a huge effect on Transience estimates and I did not expect that, given that Survival is conditional on Transience and not viceversa (so I am a bit confused about that).
These are the models structures of the lowest AIC models (but DeltaAIC>5), they are identical except for Survival parameterization:
i. {IS(i) T(sex+dry years) S(sex*dry years) C(sex*time) SEX(time) VIS(time) CORR(sex+Trend)}
ii. {IS(i) T(sex+dry years) S(juvenile*sex+adult*sex*dry years) C(sex*time) SEX(time) VIS(time) CORR(sex+Trend)}
The GEMACO sentence for Transience step is: a(1).[f+t(1 7, 2 3 4 5 6 8 9 10 11 12 13)]+a(2) where t(1 7) represent the dry years, a(2) is set to zero in the IVFV (just two age classes specified in the MODIFY button).
The first model (model i above - lowest AIC) estimates no difference between the transience probability of females and male whereas the second (model ii above) and all the rest in the set of candidate models find a quite strong difference in transience probability between females and males (females much more dispersant).
Any idea on what might be going on?
I thought about a local minima issue but it doesn't make much sense because I found the “wrong” pattern in sex propensity to transience in all the models and I have also re-run the models by increasing a lot the number of multiple random initial values and I always get the same results.
Thanks for any help,
Simone