Mixtures for multi-scale models

[s02hoff]

Hi all,

I am using multi-scale occupancy models to investigate the influence of various habitat covariates on occupancy at the Psi and Theta scales using RMark and have a question regarding the coding for my sampling scheme. There are 50 grid cells at the Psi level, and within each cell, there are 3 bat detector locations (Theta scale). Each detector operated for 8 nights representing 'visits'. My question is whether to set mixtures to 8 or 3? I have tried both ways and am struggling to understand the model output.

Thanks!

[jhines]

Mixtures should be set to the number of devices, or 3 in your case.

If you want to experiment with different scenarios for the multi-scale occupancy model, there is a function in RPRESENCE package, genpresEVmm, which generates expected-value occupancy data. Here is some sample code showing how to use it...

Code: Select all

library(RPresence)
a=genpresEVmm(K=6, psi=.75, p=c(.5,.6), theta=.4, MPS=2)

library(RMark)
data=data.frame(ch=apply(a$det.data,1,paste,collapse=''), freq=a$frq)
pd=process.data(data,model='MultScalOcc',mixtures=2)
dd=make.design.data((pd))
m1=mark(pd,dd,model.parameters = list(
  Psi=list(formula=~1), Theta=list(formula=~1), p=list(formula=~time)))

[bam59]

Hello all,
I am glad this question was asked, as it often stumps me! Based on the original post's description, theta is anticipated to vary by location within the larger grid cell where the three detectors are placed (and would perhaps interpreted as habitat use). Each detector was out for 8 nights.

Here's my confusion. In Program MARK, the specifications for this model include the following, "Specify the number of surveys (L) for this data type" and "There are K theta parameters for the K primary samples." With the L and K notation, it is clear that in the case of the original post, L = 8 and K = 3 (right?). In the RMark documentation, "mixtures" is defined as: "number of secondary samples for MultScaleOcc model". Does that mean that mixtures = L? If so, it seems like mixtures would be set to 8 for the example given by the original poster.

I'd appreciate any help in getting my head around this!

Best,
Brittany

[gwhite]

Yes, L = mixtures. But don't ask me to explain L versus K!

"Specify the number of surveys (L) for this data type in the 'Surveys (L):' box."
"Specify the length of the encounter history (L*K) for K primary samples."
"There are K theta parameters for the K primary samples.", ;
"See Nichols et al. (2008) J. of Applied Ecology for definitions "
"and notation, where their example is K samples with L devices (surveys)."

[bam59]

So...in this case, is the consensus that the original poster would set mixtures = 8 or = 3? So far we have one vote for each. Anyone care to break the tie?

[jhines]

I think the definitions of “samples” and “surveys” can be confusing. In Nichols 2008, K is the number of sampling occasions (8 in this case), and L is the number of sampling devices (3 in this case). There, “survey” seems to refer to any sample taken by any method (L*K).

In Mark, K is the number of repeated samples and L is the number of devices at each sample. The last line of the Info box appearing in Mark when the multi-scale occupancy model is chosen says, “… where their example is K samples with L devices (surveys)”.

In RMark, “mixtures” is used for the number of methods/devices (L).

The original post indicated “50 grid cells at the Psi level” (sites) with 8 repeat visits (samples) with 3 bat detector locations, so K=8, L=3. Since the 3 detectors are at each site, I assume they are treating the detectors as different devices. That means the theta parameter (local occupancy/use) applies to each of the 8 repeat visits (th1, th2, th3,…th8), giving probabilities of local occupancy/use at each visit.

If the original poster wanted occupancy for each detector location, then each detector should be coded as a “site” (giving 50*3=150 sites) and the standard single-season occupancy model could be used.

[s02hoff]

Thank you all for your input!

I went with the multi-scale models to account for potential spatial correlation between the 3 detector locations within each grid cell, as opposed to single season models. I am interested in local occupancy estimates for each detector location, as there may be covariates influencing occupancy at the local scale (my theta covariates cover 200m buffers surrounding each detector, as opposed to my psi level covariates that include landscape predictors for the whole grid cell), but also interested in detection rate trends (i.e. how many nights of sampling are sufficient to ensure we are 90% confident we will detect the species if they are present).
Since my landscape covariates on theta are not changing each survey night, would it be better to do as single season model and include a covariate to account for potential spatial correlation (like grid cell number) as well as the local scale landscape covariates?

[jhines]

If the 3 detector locations are all estimating occupancy for the site and it's possible for individuals to be present at > 1 location for a survey, then I think you should stick with the multi-scale model. It is not a problem to have the local occupancy parameter to be a function of a landscape covariate and/or a local-scale covarite.

Your question about detection rate trends can be addressed by simulating or generating data, based on predicted values of the occupancy and detection parameters, and adjusting the number of surveys until you reach an overall detection probability of 0.90 (p*=0.90 = 1-(1-p(1))(1-p(2))(1-p(3))). The function I mentioned earlier, genpresEVmm in RPresence, would allow you to do this.