U-Care: Group vs Pooled analyses difference

[ecar026]

Hi,
The primary focus of my work is to estimate abundance at a whale calving ground using independent estimates from genetic and photo-identification mark-recapture data. There were four dedicated winter field surveys, and the survey in the final year was longer than the previous years. This is a migratory destination, and the surveys only happen for part of the calving season due to logistical constraints, so I was hoping to use the POPAN superpopulation estimate to get abundance and was testing for transiency using U-CARE. Previous researchers have had trouble estimating residency time in certain sections of the population (non-cows), and I wanted to check this would not impact the POPAN model.
The data are very sparse, so I initially tested each dataset (photo ID and genetic) overall and then broke them down by sex for each dataset to check whether there was a group effect, as had been seen in the previous residency study. Although the overall photo ID tests show signs of transiency, the group statistics do not.
What is the more powerful test?
I cannot run POPAN or CJS models structured by sex as I get errors probably due to data sparseness ("group counts not sufficient" error and convergence issues).
The table below reports the chi squared statistic (Fisher's exact test) for each test as it is the most suitable for sparse data. If someone could give me some advice that would be brilliant.
GLOBAL 3.SR 3.SM 2.CT
Photo ID data
All 0.047869 0.009559 0.80125 0.96169
Females 0.62628 1 1 0.33
Males 0.19662 0.12736 0.80125 0.16667
Unknown 0.3337 0.10156 0.80125 0.97322
Genetic mark recapture data
All 0.99259 1 0.80125 0.61429
Females 0.55755 0.15 0.80125 1
Males 1 1 1 1

thanks
Emma

[CHOQUET]

Hello,
this is not an easy subject and perhaps some people will have a better advice than me.
When you pool, you can either
* gain power and by this way detect an effect which really exists
but that you were not able to detect before because of the lack of data.
* or introduce heterogeneity because for example detection rate are different between groups of individuals.

In the first case, an effect to reduce the overdispersion (like the transient effect in our case) should be added
in our model whereas nothing should be done in the second case.

As by pooling, only the transient effect is detected then you are more probably in the first case.
What you can do is to see in term of AIC if a model with transient is better than the CJS model.
Sincerely,
Rémi