Occupancy: Defining visits in sites of varying sizes
Posted: Thu Oct 18, 2012 3:03 pmI have wrestled with an issue re: occupancy modeling for a while in numerous studies and figured I’d ask you folks here.
We are studying the distributions of a number of species in urban parks. Often this involves occupancy/detection modeling, where we sample systematically or randomly in many different parks. I’ll explain my question using a camera trapping situation, but it could apply to point counts or other methods.
Cameras within individual parks are likely not independent, closed sites, and thus we use whole parks as our separate sites. However we are not sure how to then define visits. The first option would be to make visits be the cameras in each park.
e.g., say I deployed cameras in 4 parks for 1 month (4 weeks). My data could be arranged as:
Where a 1 would mean I detected the species at least once at that camera over the month, and a 0 would mean I didn’t detect the species at that camera ever.
The problem arises because these parks are different sizes and thus to adequately sample them, big parks get more cameras. Some of our parks get 2 cameras and some get 12 – 15 keeping a steady camera density. It seems that if I had a park with just, say, 2 cameras and did not get any detections, it could come out that the estimates for that park would be less precise since I “only sampled it twice”, even though in truth I sampled it equally – proportional to its size – as the larger parks with more visits. We would ideally like to be able to account for differences in detection rate across cameras, so I would like to go with the visits = cameras setup where I can enter sampling covariates on p if possible.
Alternately, we could define a visit as some time period (say, a trap-night or –week) in which the whole park was sampled, where a 1 or 0 would mean I either detected or didn’t detect the species anywhere in the park within a particular week. I have done analyses using visits=trap nights, but we are less interested in detection changes over time. It also would not reflect legitimate differences in sampling effort or camera density when cameras fail or are stolen.
It seems like these issues would arise in many occupancy studies where sites vary in total size (and thus in the amount of sampling effort needed). I’d appreciate anyone's thoughts or advice.
Thanks
chris
We are studying the distributions of a number of species in urban parks. Often this involves occupancy/detection modeling, where we sample systematically or randomly in many different parks. I’ll explain my question using a camera trapping situation, but it could apply to point counts or other methods.
Cameras within individual parks are likely not independent, closed sites, and thus we use whole parks as our separate sites. However we are not sure how to then define visits. The first option would be to make visits be the cameras in each park.
e.g., say I deployed cameras in 4 parks for 1 month (4 weeks). My data could be arranged as:
- Code: Select all
Cam1 Cam2 Cam3
Park1 0 1 1
Park2 1 1 1
Park3 0 0 0
Where a 1 would mean I detected the species at least once at that camera over the month, and a 0 would mean I didn’t detect the species at that camera ever.
The problem arises because these parks are different sizes and thus to adequately sample them, big parks get more cameras. Some of our parks get 2 cameras and some get 12 – 15 keeping a steady camera density. It seems that if I had a park with just, say, 2 cameras and did not get any detections, it could come out that the estimates for that park would be less precise since I “only sampled it twice”, even though in truth I sampled it equally – proportional to its size – as the larger parks with more visits. We would ideally like to be able to account for differences in detection rate across cameras, so I would like to go with the visits = cameras setup where I can enter sampling covariates on p if possible.
Alternately, we could define a visit as some time period (say, a trap-night or –week) in which the whole park was sampled, where a 1 or 0 would mean I either detected or didn’t detect the species anywhere in the park within a particular week. I have done analyses using visits=trap nights, but we are less interested in detection changes over time. It also would not reflect legitimate differences in sampling effort or camera density when cameras fail or are stolen.
It seems like these issues would arise in many occupancy studies where sites vary in total size (and thus in the amount of sampling effort needed). I’d appreciate anyone's thoughts or advice.
Thanks
chris