Binomial Mark Resight and Camera Trap Data

[amurphy]

Hello all,

I had a question regarding the use of binomial mark resight/spatial mark resight models for the analysis of camera trap data. I searched the forum but I couldn't find anything specifically on this issue...I'm sorry if I did miss an explanation! I'm a graduate student in possession of a good-sized camera trap data set with multiple captures of fossa (Cryptoprocta ferox) and I'm attempting to estimate their density, as well as compare density estimates between mark resight and spatial mark resight models. However, my sample sizes are small, so I'm trying to use Bayesian inference.

I know that Royle et al. 2009b (Bayesian inference in camera trapping studies for a class of spatial capture recapture models) and Royle and Gardner (2011)implied that multiple detections of one animal at one trap during one short sampling occasion (say a trap night) might not be independent or provide much data, so one can use a binomial/Bernoulli encounter model as a "reduced-information" version of the Poisson encounter model. I just want to make sure that this is indeed correct and scientifically sound (it certainly seems logical) and that I can use the binomial instead of the Poisson.

Another question: I know that Chandler and Royle (2011) worked on a spatial mark resight model using a Poisson encounter model. They stated in the discussion (under Alternative Observation Models) that they were currently looking into a Bernoulli/binomial observation model. Does anyone have any further information about the progress of this model? I don't know too much about modelling so I'm hesitant to go and substitute binomial for Poisson without having someone more experienced make the first step.

[bmcclintock]

When using camera traps, care must be taken with "reduced-information version" mark-resight analyses because while it is usually easy to keep track of multiple detections of marked individuals within a sampling occasion, it is usually difficult to determine whether an unmarked individual has been detected multiple times (e.g., did I just detect the same unmarked individual twice, or was it two unmarked individuals detected once each?). Failure to correctly reduce the unmarked detections will induce bias in abundance (or density) estimators. The rarity and behavior of the species may help you assess whether you can reliably reduce the unmarked detections.

If you deployed telemetry collars on your marked individuals, these might be worth checking out:

Sollmann, Rahel, Beth Gardner, Arielle W. Parsons, Jessica J. Stocking, Brett T. McClintock, Theodore R. Simons, Kenneth H. Pollock, and Allan O'Connell. In press. A spatial mark-resight model augmented with telemetry data. Ecology. http://dx.doi.org/10.1890/12-1256.1

Ivan, Jacob, Gary C. White, and Tanya M. Shenk. In press. USING SIMULATION TO COMPARE METHODS FOR ESTIMATING DENSITY FROM CAPTURE-RECAPTURE DATA. Ecology. http://dx.doi.org/10.1890/12-0102.1

Ivan, Jacob, Gary C. White, and Tanya M. Shenk. In press. USING AUXILIARY TELEMETRY INFORMATION TO ESTIMATE ANIMAL DENSITY FROM CAPTURE-RECAPTURE DATA. Ecology. http://dx.doi.org/10.1890/12-0101.1

[murray.efford]

Brett's advice is good. I would add -

...my sample sizes are small, so I'm trying to use Bayesian inference.

Superiority of Bayesian inference with small sample sizes is a line put about to sell Bayesian methods, not a scientific fact, despite its embarassing recitation in several recent papers. There are better reasons for using Bayesian methods, and for avoiding them.

binomial/Bernoulli encounter model

A binomial model and a Bernoulli model are different beasts. In a binomial model there are multiple trials (e.g. on each day of a week) and independence is still assumed among the trials. In a Bernoulli model there is only one trial, with a binary outcome. Both possibilities are available for capture-recapture in 'secr', so it's worth getting this straight. Of course you can see a Bernoulli trial as binomial with a size of one, but that's not very helpful. I'm still looking for a strong spatial mark-resight dataset before wheeling out mark-resight in 'secr', so I guess this doesn't help you directly.

Murray

[amurphy]

Thank you for the replies, they are much appreciated. Sorry for the binomial/Bernoulli mix-up; too many papers, not enough statistical background to make sense of them all!