Multi method and multi season/single season approach

[Catia]

I’m trying to estimate occupancy for a large carnivore. I have data from 2 different years, obtained from 2 two survey methods: Sign surveys along transects inside cells and cameras placed in the same cell grids. Here I have two problems related to the sampling design, the large home range of the species and the small size of cells:

A - The assumption of enclosure is not respected. I read that if I use the term “use” instead of occupancy that should not be a problem. Am I right?  

B - The distance used between cells is not enough to avoid special auto correlation in both methods. Will that be a problem?

 

 

My objectives:

Know which covariates are important to occupancy

Know if there is any difference between year 1 and year 2.

Know if the occupancy and detection estimates are different between methods and if I use the two methods combined

 

Being so, I want to know if:

1 -I could use a multi-method model and create models with and without “year” as a covariate on occupancy and then choose the best model? If the best models are the ones with “year” as a covariate does that means that there are differences between years? If it is and I want to know the differences in occupancy between the two years, shouldn’t I do a separate model for each year?

Or

2 - I should do Multi-season models for each method in separate and compare the differences between them.

Or

3 - I should fuse the 2 approaches, if it is possible.

 

I have another question related to seasons:

4 - As I’m using data for an entire year and I know that seasons such as Winter and Summer will probably affect the behavior and distribution of the species along the study site, should I use “season” as a covariate too in a single season model?

 

And another question:

5 – May I use the probability of detection history calculated for each sampling site as an approach to the “occupancy” in each sampling unit?

Thank you!

[John Clare]

Hi Catia,

I'll give it a shot, but nearly all of these questions do not have a quick answer. I think there are other threads related to spatial autocorrelation and multi-method models on here, so it might be helpful to search around.

A - The assumption of enclosure is not respected. I read that if I use the term “use” instead of occupancy that should not be a problem. Am I right?  

MacKenzie and Royle (2005) first propose this; anecdotally, these words seems to be used interchangeably in inconsistent fashion. You might take a look at Efford and Dawson 2012 (Ecosphere), Latif et al. 2015 (JWM), etc. for more detail. This assumption is essentially always violated for for studies using point detectors (or small radius point counts).

 

B - The distance used between cells is not enough to avoid special auto correlation in both methods. Will that be a problem?

 

Maybe? If the sites aren't independent, estimates can be overprecise. Whether psi/z exhibit autocorrelation is difficult to evaluate without fitting a model that explicitly incorporates some sort of spatial term--there are many different ways to approach this, and maybe somebody else can suggest a specific approach (note, unmarked does not allow these models).  Because environmental predictors also tend to be autocorrelated, it is a) not always clear whether an autocorrelative pattern relates to the environment (a missing covariate, maybe), or something extrinsic to the environment; b) autocorrelated environmental predictors may not be estimated very usefully if there is an explicit spatial structure. I can't remember the name (stocc? spocc?), but Devin Johnson put out a package that allows one to compare (closed, single-method) occupancy models that incorporate a couple spatial formulations vs. the standard model--if you want to go down this road, might be a place to start. 

 

1 -I could use a multi-method model and create models with and without “year” as a covariate on occupancy and then choose the best model? If the best models are the ones with “year” as a covariate does that means that there are differences between years? If it is and I want to know the differences in occupancy between the two years, shouldn’t I do a separate model for each year?

A model with a year coefficient for psi and separate year coefficient and intercepts associated with each method's p is exactly the same as a separate model for each year. If you use the first approach, you do have the capability to model certain parameters as constant (say, p for each method). You could also fit the normal single-season model with a year term (or multi-season model), but use a dummy covariate to distinguish camera sampling intervals from transects.

 

Or

2 - I should do Multi-season models for each method in separate and compare the differences between them.

Yeah, in principal, occupancy is scale dependent, so it should be different for the transects and the camera (the transects cover more space and are more likely to intersect with animal space use), and this approach in some ways makes the most conceptual sense--the methods sample different things. But I don't know that you get anything useful out of fitting the models and finding, for example, that occupancy is greater within the space sampled by one method than another.  

 

Or

3 - I should fuse the 2 approaches, if it is possible.

I think you can fit a multi-season multi-method model in MARK (see robust design multiscale occupancy; whether you fix availability or estimate it is another consideration). This option or number 1 make the most sense to me.

 

I have another question related to seasons:

4 - As I’m using data for an entire year and I know that seasons such as Winter and Summer will probably affect the behavior and distribution of the species along the study site, should I use “season” as a covariate too in a single season model?

Maybe A) refers to closure in the sense that animals are moving at a much larger scale (is this continuous 2 year sampling?)? If so, the issue of closure is a little trickier.   

 

And another question:

5 – May I use the probability of detection history calculated for each sampling site as an approach to the “occupancy” in each sampling unit?

Finite sample occupancy (called via ranef in unmarked, or reported as occupancy|y in presence) is the best guess at whether a specific site is occupied or not (see unmarked's documentation for ranef for a brief description, or MacKenzie's 2005 paper in JWM for a more detailed description)...so...I guess I don't know what you mean by approach, but that's what the quantity is.

John