Strong understimation of population size

[Granjon]

I am running SECR models to estimate the number of chimpanzees in a population from opportunistically collected fecal samples (polygon search) in Uganda in an area of about 50km2.

My point is actually to compare the estimates with the real number of individuals, so I am investigating only 1 community of chimpanzees whose size I know, and the sampling area is only a little bit bigger than their territory.
I ran different SECR models, and multiplied the density by the search area to obtain population size. First I pooled both sexes together, and then ordered sex into Sessions (S). Here is a summary of the results:

Model pop.estimate lowCI highCI
D-1,g0-1 86.78 64.51 116.74
D-1,g0-h2 189.76 136.08 264.61
D-S,g0-1 86.80 60.09 125.38
D-S,g0-h2 189.76 127.53 282.37
D-S,g0-S 92.63 60.63 141.67
D~S, g0~1, sigma~S 82.16 56.21 120.08
D~S, g0~S, sigma~S 93.96 60.09 147.05



The true population size is 189, and I am surprised at the strong underestimation of the estimate from the null model, which is the one most supported by AIC (67%). Even if we only consider the population of adolescent and adult individuals, which represent 90% of the samples, the population size would be 128, and still not included in the confidence interval.

The model with g0~h2 does provide a good estimate, but its confidence interval is so large that it does not seem very informative to me.

I am trying to account for these results that do not really speak for SECR in my case of polygon search. Can the estimates be affected by the choice of my searching area, given that it was based on the community's territory?
I had considered that my capture data was reasonable (137 samples, 81 individuals, about half of them captured twice or more, 10 of them capture 3 times or more), was I right to assume so?



I was actually a bit disappointed by the results, and was wondering whether there was any obvious reason that I might have missed.

Thanks
Celine

[Bryan Hamilton]

Small mammals certainly are not chimpanzees but I'll share a little of my experience. I had trouble with this with some small mammal trapping data. This is the answer I got from Murray. I'm not sure if its relevant in your case but it might have to do with the way you've set up your sampling units (traps in my case.

I think using groups to get grid-specific density estimates is likely to fail. Groups are for permanent population classes such as males and females. When you fit a grouped model you mean that all individuals in each group follow the overall population model i.e. they are located at random across the region. However by defining group by location you deliberately violate this assumption: all animals in one group live in a selected fraction of the landscape, and density is inferred to be mysteriously low (because of the absence of animals from this group on other grids). This would not apply if you aligned all the grids to have the same coordinates and trap numbering - then a grouped analysis might work as you expect. An individual that moved between grids would then just just show as an odd within-grid movement.

The better way to get grid-specific estimates when between-grid movements are rare is to define grids as sessions. There is no need to reassign the ID of an individual caught on two grids as each grid is separate in the analysis.

Once I corrected this, my density results became much more reasonable. I'm pretty sure N is based on an extrapolation of density in 'secr'.One consistent result I've seen in my data is that density estimates in 'secr' have been consistently lower than my count data/sampling area. This is another consideration for you. It wasn't intuitive for me, why this was the case.

[Bryan Hamilton]

Another paper that might be relevant to you:

Gerber, B. D., and R. R. Parmenter 2014. Spatial capture–recapture model performance with known small-mammal densities. Ecological Applications 25:695-705.

They showed that SECR estimates ( believe in a bayesian framework) were biased low with know small mammal densities.

[murray.efford]

Hi Celine
Something is obviously going wrong, and I'm reluctant to believe it's an inherent problem in SECR as implied by the Gerber and Parmenter paper - with some forensic work we can often locate the problem elsewhere, e.g. in the way the 'tests' were conducted (I have ongoing discussions with Brian Gerber and Jake Ivan on this). Which of course is what you are asking for. I would like to understand more about your chimp study, but have limited time in the next few days. Will get to this when I can, and may ask for more detail.
Murray

[Granjon]

Thank you Bryan and Murray for your input.

I was a bit reluctant in posting all the details in an open forum, but I would be happy to share them with you per e-mail whenever you have time.

Céline

[murray.efford]

I understand. Do send me data offline if you like - you should find my email address easily.
Murray