[tlocoh-info] Parameter selection in a linear moving species

[Andy Lyons]

Hi Irene,

Sounds like an interesting study. A few comments below:

On 6/11/2014 9:43 AM, Irene Weinberger wrote:
> Dear tlocoh-Group
>
> I have followed the discussion with Anna and Thomas and it has clarified
> some questions. Unfortunately not all. I am working with otters that
> move along rivers and I think that LocoH's are a fine approach to
> estimate home range size as they are built nicely along the linear
> features. However, I am facing few problems and I wonder what you think
> about it:
>
> a) I have data of only few animals but from them, I have a time span of
> radiotracking of 6 months to about 28 months. Following your advices to
> Thomas, I used an individuell varying s for each individual for the
> timespan of 6 months (as this is the shortest tracking period) using the
> "sfinder". So far, I did not subsample my data (I do have busts of
> locations of 2-16 hours every 15 min followed by days when the animal
> has not been tracked or just single locations within those days and all
> varies among the individuals).
> I was wondering if I need to subsample of if I account for this when
> using s > 0.

In general, incorporating time into nearest neighbor selection (s>0) 
will have the effect of treating bursts of locations separately. Two 
points that are close together in space will nevertheless be considered 
very far apart (for determining nearest neighbors) if they are widely 
separated in time (relative to other observations). So presuming you 
want each burst of locations to be treated separately, if you use s>0 
there shouldn't be a need to subsample your data. That being said, it is 
also worth noting that if you select nearest neighbors using the 
k-method, and you have one or more 'burst' of points with fewer than k 
locations, then by necessity those hulls will be constructed by 
locations taken from other bursts as well (which may be undesirable). 
For example if you have a burst of 10 locations (& no other locations 
for days before / after), and you are using a k=15, then constructing 
those hulls will necessarily require using points that were days apart. 
If that's undesirable, you could use a smaller value of k or the 'a' 
method (preferable).

Another issue to keep in mind when you have irregular data sampling is 
that the density isopleths (which essentially mirror the density of 
points) could in fact be a reflection of the sampling as opposed to how 
the individual used space. In other words, an area where the individual 
was located during a period of heavy sampling may appear as a 'core' 
area (lower isopleth level) relative to an area the animal happened to 
be when sampling was light, due purely to the sampling. Whether this is 
an issue or not will depend on your data, your questions, and the animal.

Subsetting data in different individuals to a achieve a comparable 
sampling interval is most essential when you are computing hull-based 
association metrics, but that doesn't seem to be the case in what you 
are doing.

> b) My animals belong to a linear moving species that keeps to rivers and
> streams. I understand that I should be using the same k for all
> individuals for the estimation of the home range. However, while most of
> the LocoH's converge at around k=38, I have two individuals where the
> locoH's converge only at high ks (85 and 190, respectively) and if I use
> the highest k for all of the individuals, then I get much too large home
> ranges. I am here at a loss how to proceed and would like to hear if you
> have any suggestions.

For exactly the reasons you state, I don't fully subscribe to the 
principal that one should always use the same value of 'k' (or 'a') for 
different individuals (or the same individual during different time 
periods). Using the same value of k makes sense for different 
individuals when the movement data are comparable, i.e. similar sampling 
rate, similar proportions of time spent in different movement modes, 
etc. To illustrate this point with a theoretical counter-example, if you 
have one individual who is territorial (protecting a nest for example), 
and another individual who wonders around a lot, there would be no 
reason to expect the same value of k would best represent space use for 
both individuals. Another fabricated example would be if you have one 
individual tracked with the sampling rate of 15 minutes, and a second 
individual with 1 hour sampling, you would almost definitely need 
different values of 'k' to achieve a comparable balance between Type I 
and Type II errors (or what I call 'Swiss cheese' territories and 
spurious cross-overs). I would argue that the best value of k should be 
the one which produces the most authentic model of space use, which for 
most studies includes 1) differentiating the gradient of intensity of 
space-use (or another hull metric depending on the question), 2) 
delineating spatial boundaries (hard edges) in movement, and potentially 
3) delineating temporal partitions of space use. I recognize however 
that selecting parameter values for each individual separately, based on 
a common set of principles, can be more work, and for large numbers of 
individuals may not even be feasible. I invite other comments.

I don't know if that helps or makes your choices more confusing. I also 
might have misconstrued your question - are you seeking to construct 
home ranges for each individual separately, or perhaps estimate a single 
UD for all individuals combined?

Andy


> Best wishes
> Irene
>
> _______________________________________________
> Tlocoh-info mailing list
> Tlocoh-info at lists.r-forge.r-project.org
> http://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/tlocoh-info
>