www.phidot.org

Hello,

I have a really nice DNA-based capture-recapture data set that we collected on brown bears this past summer that I am anxious to analyze using secr. However, unlike other data sets that I've analyzed, the sampling method that was used is making it problematic for me to define discrete sampling occasions for the study area as a whole.

First, a quick description of our data. We deployed almost 550 detectors, including single-catch hair snares and scent-baited corrals and rub trees, over a study area of approximately 1700 square km. In addition, we had 18 collared bears that provided location information at 20-30 minute intervals during the hair sampling session, as well as a couple dozen camera traps in which uniquely identifiable bears were captured. Sampling took place over the course of 6 weeks and we were able to positively identify almost 180 individuals from the 450 DNA samples that successfully amplified, with between 1-11 detections per individual. In short, we have what seems to be a very robust data set.

I am struggling with how to divide the sampling effort into discrete sampling occasions that can be applied to every detector within the entire study area. The detectors were visited every 7-10 days to collect hair samples and to reset the single-catch snares. Collection began at one end of the study area and continued towards the other end until all detectors were checked. Due to the long distances and rough terrain, this was not a fast process, so by the time the last detector was checked, it was time to start all over again with the first detector. Thus, sampling took place on a temporally continuous basis over a rolling spatial window. Does that description make sense? It would be easy to define the sampling occasions for individual detectors, or even groups of detectors, but I'm not sure there is a way to specify this arrangement in the trapfile.

Do you have any suggestions or recommendations?

Hi

It does sound like you have rich dataset! As my first stab at a answer...

It doesn't sound as if the temporal dimension is important in itself (6 weeks isn't long) unless perhaps there are strong learned responses. If you don't care much about time then the temporal relation between samples does not figure in the model and you can just equate 'occasion' with 'check'. The asynchronous sampling example in our 2013 paper (MEE 4:629-636) may help you think about this.

You also have the option of collapsing the data to a single 'occasion' (using CH1 <- reduce(CH4, outputdetector = 'count', by='ALL'). If each bear can only be recorded once per site per check (binary proximity data) then the aggregated data are binomial counts with binomial 'size' equal to the number of checks. You can specify either binomN=4 (assuming 4 checks at each site) or binomN=1 (relying on the 'usage' attribute to convey the site-specific binomial size).

If you do have strong learned responses (presumably to lure/bait) then you would need to keep track of time, but it gets messy. Perhaps you could slice time into weekly units or smaller and in each one for each site account for the checks that took place and the time elapsed since the last check... I hope you don't have to do this!

Murray

I suspect that I will have a strong behavioral response to the scent lures. We've seen similar behavior with wolves when scent-baited hair snares are used--camera data show that they sometimes come back several months after the snares have been removed and roll on the exact spot of ground where the snare was once located. Thus, I think the approach of slicing the session up into weeks and then including non-binary usage data for each detector to account for the amount of time between checks might be the way to go. In theory it should be simple to set that up--just rather time consuming.

I would also like to incorporate the camera detections, as it provides additional information, especially when an animal leaves hair but the DNA does not amplify. Is it possible to include multiple trap types in the same model? I suppose all of my detectors could/would get modeled as proximity detectors, even though some of them are strictly single-catch?

Thank you for the paper reference. It should prove useful.

Bravery in the face of complex and messy analyses is commendable, up to a point! I am still not sure it's worth it. If the learned response is to a particular site (as in your wolf example) then you would model it as a 'bk' effect in secr. In my limited experience, ignoring this sort of learned response (specifying a null model when it should be ~bk) is far less damaging in terms of bias than ignoring a global learned response (~b).

I am also wondering whether it isn't perfectly OK to fit a ~bk model to asynchronous data (occasions defined by check number) because what matters is the history of visitation to a particular site: when you re-check a site you _do_ know exactly the bears that have previously visited that site, and these will be modelled with a different (higher) detection probability at that site (for the model, we don't care about what happened at other sites). So I suggest you go ahead with bk models (a priori these seem more likely than a global response, and I have certainly seen cases where ~bk fits better to brown bear data than ~b).

How were the bears identified at cameras? From natural marks or artificial marks? It is possible in principle to combine these data. That's not to say it's worth it. There are issues to do with (i) how the camera 'sample' relates to the population as a whole, and (ii) whether uncertainty about sigma is really limiting (given that sigma and perhaps D are the only parameters shared between the two parts of the model).

Murray