[AniMov] kernel size?

[Clément Calenge]

Hi Paolo,

>We did some kernel analyses (h=href); everything seems to run fine, but the 
>resulting ranges seem to me a bit exaggerated (see 
>http://www.faunalia.it/download/kernel_9.png). Of course, the right thing to 
>do is to compare R results with those of other programs, but did someone 
>already tested this?
>  
>

Actually, "h=href" corresponds to a smoothing parameter estimated under 
the hypothesis that the underlying utilisation distribution is normally 
distributed (UD unimodal, i.e. one center of activity, and symmetrical). 
This is not the case in your example. When the UD is not bivariate 
normal, the use of href greatly overestimates the UD (see Silverman 
1986, Worton 1995 J. Wild. Manage.).

I never compared the estimates of adehabitat with other softwares, when 
smoothing values are estimated with the ad hoc method. But before 
developing adehabitat, I was using the software RANGES V, which also 
returned greatly overestimated home-ranges with "multi-center" home 
ranges. So that in my opinion, the reference method would return similar 
results whatever the software (though not exactly identical, because the 
different softwares do not use exactly the same algorithms - not the 
same size of the grid, not the same way to compute the limits of the HR 
from the UD).
I agree that it would be interesting to perform such a comparison...

One alternative, when several centers of activity are present, is to use 
the LSCV method, but, again, the results would differ among softwares 
(and more dramatically). For example, I used the dataset "puechabon" of 
the package adehabitat (all animals pooled), and I estimated the UD with 
different home range estimation programs. The smoothing values (in 
metres) returned by these programs are:

83 m        adehabitat
310 m      Arcview - animal movement analysis (AAMA)
44 m        ranges V
802 m      Calhome

And two other programs use one smoothing parameter for x and one for y:

59 m for X and 161 m for Y    The home ranger
131 m for X and 364 m for Y   kernelHR

As you can see, there is much variation in the estimation. The main 
cause of variation is that the different softwares do not use the same 
algorithms to smooth the UD. In addition, the algorithm often fail to 
minimise the smoothing error, so that bad results are returned by the 
function  (this is a property of the method, see the help page).

Finally the last method (which I prefer, personnally), is to specify a 
value for the smoothing parameter (the same for all animals), based on 
some visual exploration of the data.
HTH,

Clem.

-- 
Clément CALENGE
LBBE - UMR CNRS 5558 - Université 
Claude Bernard Lyon 1 - FRANCE
tel. (+33) 04.72.43.27.57
fax. (+33) 04.72.43.13.88