Using Integrated Multi-Species Occupancy Models (intMsPGOcc) with Camera Trap and AudioMoth Data

[José Wagner Ribeiro Júnior]

Dear all,

I am currently working with a dataset collected over three months in 2022 using both camera traps and AudioMoth recording devices (ARDs). The study was conducted across continuous mature forests, secondary forests, and forest fragments of 1, 10, and 100 hectares. The sampling design involved 10-hectare units where both ARDs and camera traps were deployed.

In the 10-hectare and 100-hectare fragments, as well as in secondary and continuous forests, five camera traps were placed: one near the center and four in the corners, with a 50-meter buffer from the fragment edges, covering the 10-hectare sampling unit. Additionally, two ARDs were placed at least 100 meters apart and 100 meters from the edges, adjacent to two of the cameras. In the 1-hectare fragments, two cameras were placed at opposite corners to maximize distance, along with one ARD.

I plan to use the Integrated Multi-Species Occupancy Model (intMsPGOcc) to account for detection-nondetection data from both ARDs and cameras, to estimate species occurrence. However, I am facing challenges regarding how best to handle the proximity of some of the recorders and cameras, especially since the model currently lacks the functionality to directly account for potential spatial correlation between devices.

One approach I am considering is treating ARDs and cameras as separate sampling units in the analysis and 10-hectare sampling units as random effects to account for spatial clustering, given that each 10-hectare unit contains 3 to 7 devices. Would this approach be appropriate? For example, occ.formula = ~ fhd + (1 | 10ha_sampling_unit)

Any insights or suggestions on how to best approach this issue, especially in terms of combining data from two detection sources in a multi-species occupancy model, would be greatly appreciated.

Best regards,
José Ribeiro

--

José Wagner Ribeiro Jr (Xuleta), PhD         

Quantitative Ecologist

CV Lattes

Researchgate

Twitter

[gilles colling]

Dear José,

I think I might have some input on your approach. To me, it seems reasonable to treat ARDs and camera traps as separate sampling units since their detection probabilities are likely different. Incorporating random effects for both the 10-hectare sampling units and device types seems like the correct approach.

A random effect could particularly useful because it accounts for variation across different levels of a factor, like your 10-hectare sampling units and device types, without having to estimate a separate parameter for each. In your case, devices within the same 10-hectare unit may be correlated due to similar environmental conditions or spatial proximity. By including a random effect for these units, you can account for this clustering without assuming the same occupancy or detection probability across all units.

Similarly, different detection devices (ARDs vs. camera traps) likely have different detection probabilities. Adding a random effect for device_type allows you to account for that variability in detection probabilities across devices while still sharing information across them.

A simple random intercept model would account for baseline differences in detection or occupancy between units and devices:

occ.formula = ~ fhd + (1 | 10ha_sampling_unit) + (1 | device_type)

If you suspect that the covariates, such as forest height density (fhd), might have different effects across units or devices, a random slope model would allow the effect of fhd to vary across them:

occ.formula = ~ fhd + (1 + fhd | 10ha_sampling_unit) + (1 + fhd | device_type)

Random effects are great for addressing this kind of variation in a flexible way. Testing both models should help you determine which best fits your data.

Let me know if you’d like to discuss further!

Best regards,

Gilles Colling

Sent from my iPhone

On 16.09.2024, at 18:37, José Wagner Ribeiro Júnior <jwribei...@gmail.com> wrote:



--
You received this message because you are subscribed to the Google Groups "spOccupancy and spAbundance users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to spocc-spabund-u...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/spocc-spabund-users/CAPAy4WyoMrEjDYC2H3wifD937Mxz8zqdyau%3DD3s9xO4_H3zPvQ%40mail.gmail.com.

[gilles colling]

Dear José,

I mean ofc to include the random effect for the device type in the detection part not the occupancy part.

det.formula <- ~ (1 | device_type)

On 16.09.2024, at 19:11, gilles colling <gilles.c...@gmail.com> wrote:



[José Wagner Ribeiro Júnior]

Dear Gilles,

Thank you for your helpful insights and for taking the time to respond. Your suggestions have provided the clarity I needed, and I feel confident moving forward with the approach you outlined.

I really appreciate your offer to discuss further, but for now, the information you've shared has been sufficient.

Thanks again for your support!

Best regards,
José Ribeiro

[Jeffrey Doser]

Hi José,

I just wanted to add a couple of thoughts/comments related to Gilles' great advice:

1. With the intMsPGOcc() function, random effects are only supported for the occupancy portion of the model, so you won't be able to include any random effects in det.formula. Further, spOccupancy currently only supports random intercepts, so if you wanted to explore a random slope you won't be able to do that. It's been on my todo list for a while to add in functionality for random slopes, but I don't have a great sense of when I might get around to adding that in.
   
2. I am close to getting an update to the package on CRAN that would allow for random intercepts on both occupancy and detection in single-species integrated occupancy models (both intPGOcc() and spIntPGOcc()). That should be on CRAN within the next month or so, hopefully sooner. I'll send another email on this thread when there is a stable version on the GitHub development version that you could explore as well in case it takes a while to get onto CRAN. If you need random effects on detection probability, that could be an option to potentially use (of course would have to do things on an individual species basis though, which could limit the number of species you are able to model).

Cheers,

Jeff