Dependent captures and differences in sampling effort

[Carlotta Gelsomini]

Hello everyone,
Sorry in advance for the rather long message. I have several related questions, and I’d really appreciate your advice.
Our camera-trapping design is not a regular grid, but camera traps are placed at known marking sites that males frequently visit. I just wanted to mention this for completeness.

1. Our species has two male tactics that differ in home-range size and, consequently, in capture probabilities. Within each tactic, individuals can form groups. The territorial males, with smaller home ranges and higher capture rates, are more frequently found in small groups (typically 1–3 individuals), although such grouping can occasionally occur in the other tactic as well. Because these individuals move almost exclusively together, captures within a group are clearly not independent.
Would it be more appropriate to treat these groups as units rather than as individual animals?
Since our main focus is on density estimates, I assume that this would need to be accounted for in the final estimates, as otherwise densities would be underestimated if we treat a group as a unit/single animal.
For some of these group-living males, we also have telemetry data from multiple individuals within the same group. 
What would be the best approach for handling this kind of structure and data in oSCR?

2. My second question concerns how much sampling effort can differ between sessions before it becomes problematic.
For example, if in one year the effective sampling effort per area was 30 days, while in another year it was 60 days, would this be acceptable as long as the difference in effort is properly reflected in the number of occasions?

3. I’d like to analyze four surveys conducted in different years, treating years as sessions. However, the areas with the camera traps were not identical across surveys. Only two areas were sampled in all four years, while the others were included in only one to three survey years.
Would it be possible to model all four years together, using years as sessions even though the sampled areas (partly) differ between years? Or would it be better to model each year separately, and additionally run an analysis restricted to the subset of areas sampled in all four years, treating years as sessions? 

Thank you very much in advance for your time and any suggestions.

Best wishes,
Carlotta 

[Daniel Linden]

Hi Carlotta,

These are great questions but you might find more useful interaction at the HME Google group (https://groups.google.com/g/hmecology), given that they are fairly general and not specific to oSCR.  Question 1 is especially tricky and may require a specialized hierarchical model that oSCR cannot fit.

As for question 2, the difference in sampling effort is not necessarily problematic as long as you account for it in the model.  Large differences in sampling can sometimes make inferences trickier if there is increased uncertainty.

As for question 3, it really depends on how much sharing of information you want across the sessions.  If all parameters are going to be estimated as session-specific, then grouping them is not needed.  But if certain sampling parameters were to be shared, which can increase precision, you might want to go through the effort of combining.  You could have 1 state space that covers all trapping areas and simply use trap operation to make clear that no data were collected in some areas during some years.  This may be more computationally expensive, depending on how many traps and state space pixels you have, but potentially worth it.

[Carlotta Gelsomini]

Hi Daniel,

thanks for your feedback and for pointing to the HME group, I will follow up there as well. In the meantime, we’ve been considering a potential approach for our group-living males and wanted your opinion. The idea is to treat each group as a single unit in the SCR model, estimate unit density as usual, and then multiply the estimated density by the average group size, including units that represent single individuals, to obtain the density of individuals. Most groups are very small (usually 2 individuals, occasionally 3), so any bias from variable group sizes would likely be limited. Do you think this could be a reasonable approach in oSCR, or are there important considerations we might be missing?

I also have another question regarding camera setup: we set up two cameras per site but record only one coordinate per site. How can I account for sampling effort if one of the cameras fails during part of the sampling period? In practice, the effect of a camera failing is likely small, at least over short periods, since animals are rarely detected by only one camera, with single-camera photos often being of low quality anyway, and most animals close enough are still captured by the functioning camera. Would it be acceptable to essentially ignore short periods when one camera is down, or is there a better way to account for this in oSCR?

Thanks again for your guidance!

Best regards,
Carlotta 

[Daniel Linden]

Hi Carlotta,

The challenge with using groups as individuals is that group size becomes a random variable that needs to be estimated.  There is no way to incorporate that estimation directly into oSCR.  This would be straightforward with a Bayesian approach and there are other more complex SCR models that deal with group dynamics.  The other potential problem with using groups as the unit is that the assumption about individual (group) identifiability being equal might get tricky.  It really depends on how groups/individuals are being identified.  Finally, group membership would also be an issue if there is a chance that it changes during sampling.

It is probably easiest to ignore detections of single cameras, especially if they are rare.  Then your sampling is defined by successful operation of 2 cameras at a site, which is a common strategy for situations where 2 sides of an animal are needed for certain identity.  If you had more single camera sites and were throwing away a lot of data, you might be motivated to model that process.  It is certainly better to have a solid design when possible.

[Rubén Portas]

Good evening!

I am running a multisession model with telemetry data for 6 out of the 7 sessions. This is data from cheetahs in Namibia. Each session is located in a different geographical country-wide location.

I am facing the following problem when fitting the oSCR models:

Error in Ytels[i, , drop = F] : subscript out of bounds
This error occurs when I include a session that has no telemetered individuals. The model runs without this error when I remove that session, which makes me think this may be related to how oSCR handles sessions with zero telemetry.

Warning (after removing the no-telemetry session):

Warning message:
Something went wrong! Try better starting values.

I’ve tried providing explicit starting values for p0, sigma, and D (and also testing models without asu), but I’m still running into optimization failures.

Any thoughts or suggestions?

.Thank you and best wishes,

Ruben

Department of Evolutionary Ecology
Leibniz Institute for Zoo and Wildlife Research (IZW)
in the Forschungsverbund Berlin e.V.

:: Evolutionary wildlife research for conservation ::

 

Projects:

GAIA Initiative    Cheetah Research Project   UL Wildlife Ecology Lab

 

Member of the IUCN Vulture Specialist Group

My Researchgate

[Daniel Linden]

Hi Ruben,

Are the telemetered individuals also captured?  One trick to accommodate the unbalanced data would be to specify a dummy individual with a single fix location in that session without telemetry.  A single location will not affect the likelihood but the setup would accommodate the structure that the code otherwise expects.

As far as the sessions that are working, an important step is to check the distribution of fixes for your telemetered individuals to make sure there are no egregious outliers.  This is the primary reason that folks attempting these methods have had problems.  Whether a location is an outlier (or egregiou outlier) is subjective, but your starting values for sigma will affect that.  In some cases, movements observed with telemetry will simply not fit the assumptions of the bivariate normal.  You can have some pretty funky patterns that still work, but others will cause problems.  So I suggest plotting the data as a sanity check, especially if those telemetered individuals were also observed at traps.

Happy to look at your script/data if necessary.

--
You received this message because you are subscribed to the Google Groups "oSCR" group.
To unsubscribe from this group and stop receiving emails from it, send an email to oscr_package...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/oscr_package/CAOcMG%3DE213T60ojuf3gsTkQAQ96nnSQDzO-wABYnHRX8PTBZTA%40mail.gmail.com.