I'm sorry to only have a question in return. Doesn't it depend on what type of question you want to answer?

You are right.

We want to determine the availability of food ressource in each deer home range.

More specifically we want to see if the availabitlity/quality of food ressources is the same in all the home ranges or if some deer have acess to more or better food ressources than others, so we want to compare the data between home ranges(deer).

I don't know if this is useful, but based on gut feeling I would think that sample locations within home ranges are per definition non-independent, because a deer would base its movements on what it finds in some place. Have a look in the GUD literature. What about a regular sampling grid within a home range? Do range size differ a lot? How static are these home ranges?

What are your main questions? Often when accounting for autocorrelation important behavioral or fine-scale data can be lost. If this involved in your main question, you might not want to account for it. I suggest reading Cushman et al. 2005 - http://onlinelibrary.wiley.com/doi/10.1111/j.0030-1299.2005.13538.x/abstract

If you are comparing conditions between homeranges, why does spatial autocorrelation matter? Perhaps I am misunderstanding but the way Amélie elaborated on the question, it sounds like you are comparing between polygons (or Kernel Density Estimates or whatever homerange estimator you chose).

Thank you for your answers and input!! It is very much appreciated!

Yes Peter, we will be using kernel density to determine the core area and then select a number of deer relocations and sample them. The objective being to compare the resource quantity and quality between core areas. I guess we are just wondering if we should take autocorrelation into account or not. We were just afraid that if we randomly select a number of relocations and that many of them are within a few hours of each other then perhaps our estimates would not be valid. Perhaps Bert's idea of creating a regular sampling grid within the core area would solve this.

I think random resampling of the locations should deal with the spatial autocorrelation. So, lets say you have 100 locations taken every 4 hrs. You randomly select 20 of them and study the vegetation within some area proximate to those locations. That sounds fine by me. What I don't understand is why even bother with kernel density esimate if you are going to use the actual locations for sampling. I also don't understand what you are comparing the used locaitons (where the animals were) to. Other used locations of other animals? Presumed unused locations? That could make a difference in the way you deal with the spatial autocorrelation.

I think it is a good idea to consider this spatial/temporal autocorrelation. If a large proportion of relocations of deer are within some certain area, and if you randomly sample all locations, you will have more sampling plots in that certain area. But what I am hearing is that you are interested in availability, so your sampling would be biased towards the places that the deer use a lot. Given that you may be interested in availability within the core area... You could first use a kernel density estimater to delineate a core area (perhaps 50% UD), then use the grid sampling approach mentioned above within that area to assess availability. Then the autocorrelation among locations was only used to delineate the area to sample, and then the sampling is completely uniform within that core area. Regardless, for papers: look at Dray et al. 2010 and another book chapter by Cushman: http://www.fs.fed.us/rm/pubs_other/rmrs_2010_cushman_s011.pdf. Hope this is useful...

I think Jared's point on the bias towards heavily used areas gets at the greater point that your question needs elaboration before more precise advice can be given. Are comparing used areas to unused but presumed available areas within the animal's or herd's homerange etc...? Jose' and others have pointed to the numerous ways that one can measure dependencies within a dataset but Stephanie's point still stands: If you take away spatial or temporal autocorrelation, you will almost certainly be losing biologically meaningful information There is a preoccupation with removing spatial autocorrelation in animal movement data but there is no way to do so without losing useful information unless, in the rare (or implausible) case that there is no spatial or temporal pattern at all. So, while it is good to quantify the amount and scale of autocorrelation (eg. fidelity to a winter site), I say use caution in removing all spatial or temporal relatedness among animal locations.

Thanks so much for the input! We're actually not looking at availability versus used, though we did not make this clear in our question. The context of the study is straightforward; we are examining the effect of an anthropogenic disturbance on deer, so we have decided to subdivide the study period into 3 sampling periods "before", "during" and "after" the disturbance. What we failed to mention is that our study period is in the fall when the quantity and quality of resources are most likely unstable and changing. We would like to account for this so that is why we were wondering if autocorrelation between locations mattered. We will definetly check out the readings you suggested!

Thanks everyone for the answers, I think we converted Amélie to Research Gate !

Hi, you might be interested in our paper on anthropogenic disturbance on dingoes. There is a lot of debate on autocorrelation in the literature related to resource selection function analysis, some of it might help with your question. We mention it briefly in the introduction: http://www.plosone.org/article/info%3Adoi%2F10.1371%2Fjournal.pone.0063931

In their paper Kaczensky et al. (2008, Resource selection by sympatric wild equids in the Monglian Gobi) dealt with temporal autocorrelation of GPS data in a different way: all locations separated by a time interval lower than 72h were given a weight less than 1 based on a specific power law. You may want to have a look at this. Good luck!

Hi, why don't you give a look at the Brownian Bridge approach. The time spent in an area is an important factor, so why get rid of it? I hope it could be useful..Stef

Hi again, if you want to assess the independenced of the relocation you can use the Schoener test (package adehabitat)...Stef

Thank you! I will look into this!

One simple way to reduce autocorrleation is to determine the time intervals at which successive traveled distances are no longer autocorrelated. You can see more detailes in the article below. I hope you find this information useful

Article Detecting autocorrelation problems from GPS collar data in l...

There is paper by Dean Anderson et al. who may give you some ideas on protocols and data collection on a more broader sense. The citation is:

Anderson, Dean M., Rick E. Estell, and Andres F. Cibils. "Spatiotemporal cattle data-A plea for protocol standardization." Positioning 4 (2013): 115-136.

Hope this is also useful.

Hi Amélie.

I agree with Stefano Anile, time spey in an area is important and in subsampling your data you'll lose information. Plus autocorrelation if actually full of information it would be a shame to throw it away! Instead of considering location (GPS fixes) why not look at movement i.e. at the step length, speed etc between locs?? Packages adehabitLT and adehabitatHS in R might help.

Hope this helps a bit!

Stephanie

Thank you for your input, it is very helpful!

Hi, Im doing my Post-Doc in Canada in Mark Boyces lab and have been working a little bit with this type of problems.

Correcting confidence levels to account for autocorrelation is one way to get around the problem (Nielsen 2002). Another way is to incorporate an autocorrelative structure in the RSF model, see chapter 5 and 6 in Zuur et al (2009). Including an autocorrelative structure is more important than modeling the autocorrelation perfectly (Zuur 2009). Note that there may be a potential problem with corrected confidence levels and including an autoregressive structure if the random locations are not autocorrelated as this leads to that only parts of the data tested in the model will be autocorrelated (Fieberg 2010).

Modeling the autocorrelation explicitly has an important value in understanding the behaviour of the organism, as the autocorrelation can incorporate behaviours affecting movement and space use, for example the state of the animal or the memory of the animal. Explicitly modeled autocorrelation may help us find patterns and get an idea of the behavioural mechanisms causing the patterns (Wittemeyer 2008, Boyce 2010).

Before doing any of those things I would start by including a temporal structure to the data (dawn, day, dusk, night for example) and the spatial structures of the landscape (ruggedness/elevation and habitat category), theese measures usually reduces the Autocorrelation to virtually Zero for ungulates (depending on the fix rate offcourse).

If you want to model the coises in a step selection function (Fortin 2005), A form of resource selection function with a more realistic approach to what is available for the animal, send me a PM or email me at [email protected] and you can have a look at a review paper Im working on (its in the review process for Movement Ecology at the moment) and some tips on what software to use and so on...

For a bit of an alternative perspective, I suggest reading the following:

http://dynamicecology.wordpress.com/2012/09/11/statistical-machismo/

http://dynamicecology.wordpress.com/2012/09/24/why-ols-estimator-is-an-unbiased-estimator-for-gls/

http://dynamicecology.wordpress.com/2013/10/02/autocorrelation-friend-or-foe/