overlap in non-resident animals monthly space use

[Genevieve Finerty]

I'm running a PCA on trajectories from wild lions to try to get at clusters of movement strategies as in Abrahms et al. 2017 Movement Ecology paper on movement syndromes. One of the metrics included is mean volume of intersection measured as the overlap between monthly 95% kernel density home ranges. In a first step I've used segclust2d to identify periods stationary in xy use, but I know a priori that some of these segments are not going to be exhibiting range residency. As far as I can see, this means they would not be appropriate for home range analysis -- which would make it challenging to calculate any kind of UD overlap. I just wanted to check that I wasn't missing anything? I think the concept of how much this is shifting is likely a pretty good indicator of their overall space use strategy, but am at a bit of a loss of a good alternative way to get at this -- I was thinking maybe the distance between centroids of monthly groups of points, but it doesn't feel quite there. Thanks! Gen

[Christen Fleming]

Hi Genevieve,

Thinking out loud: If one period isn't range resident, then it's range distribution will blow up and the overlap between it and an adjacent range-resident period would be low. On the other hand, the overlap between non-resident periods won't necessarily be that low, but the CIs will be wide regardless.

If you want, I could take a couple of days to implement the Mahalanobis distance with bias correction and CIs (https://en.wikipedia.org/wiki/Mahalanobis_distance). That will give you a relative distance between the centroids, but it's relative to the covariance and during non-resident periods the covariance will be estimated to very large. So the Mahalanobis distance between adjacent resident and non-resident periods would be very low. On the other hand, the Mahalanobis distancebetween adjacent non-resident periods would also be very low. I could also implement the simple Euclidean distance, which would just have wide CIs in these cases, but not do anything weird.

Best,

Chris

[Genevieve Finerty]

Hi Chris, 

Thanks for the quick response and the offer! 

A couple of thoughts off the top of my head -- first, in theory, as segclust2d should have already split the full trajectories into segments which are stationary in mean/sd XY, all the sections within these segments should be either all resident or all transient -- second, I think that could work really well, as the main criteria are that two consecutive non-resident sections should produce low values, with two adjacent resident sections producing high values. But I am also wondering if this wouldn't be a good alternative way of identifying the transition from one to another?

Genevieve

ps. sorry for double emailing, I thought hitting reply would automatically post to here, but apparently not

[Christen Fleming]

Hi Genevieve,

What do you want to identify/discriminate specifically?

Best,

Chris

[Genevieve Finerty]

Hi Chris, 

To begin with, at a basic, I'm hoping to discriminate between periods of residency and transcience -- which in theory I can do visually for some animals -- but I'm working with 300+ individuals with multiple years of data, so am looking for a more objective way to cluster different movement strategies. 

Classification using e.g. shape of NSD over time isn't really suitable for a couple of reasons, partly because over the full trajectories lions might exhibit a number of strategies, partly we don't always have the full path from e.g. immigration to settlement for dispersal, and partly because the strategies they seem to exhibit don't align neatly into predetermined shapes, e.g. they might settle very close to their natal range, after a year or so of dispersal, including multiple temporary resident periods. At least for some of the animals I know the context of, segclust2d seems to do a pretty good job of separating out sections where I know something changed -- but it's not always simple from that to work out what each segment relates to biologically-- and for quite a few, we have no context from observations.

So, the idea is to use a PCA of various different metrics to group together my segments into residents (and see whether there are clusters within animals exhibiting range residency for e.g. exploratory animals vs territorial vs CPFs) and transients (some of whom might be on directed or some might be more nomadic) -- so I was hoping that a good measure of overlap of adjacent periods of time within segments would be a good discriminator for the transient vs resident spectrum of this, then paired with other metrics eg. maximum squared displacement, residence time, time to return, average monthly distance/speed to further differentiate other aspects. But I was also hoping, using some labelled sections to look at how these metrics individually vary between groups.

Sorry, that's a lot of info at once -- hopefully, it's coherent enough to answer your question?

Gen

--
You received this message because you are subscribed to the Google Groups "ctmm R user group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ctmm-user+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/ctmm-user/ad3b3977-4dcf-4d4c-b370-26e4258581c1o%40googlegroups.com.

[Christen Fleming]

Hi Genevieve,

Believe it or not, that kind of classification is the second project on my table right now. My first project is on meta-analysis of home-range areas, and then my second project is on clustering-analysis of home-range areas versus dispersal (or whatever else) areas. The code for the meta-analysis seems to be working well and is on GitHub ( help('meta') ), but I have to rewrite the clustering-analysis code and I don't have a time frame for that at the moment. I might not get to it until after TWS.

I had went ahead and implemented some distance functions on the development branch on GitHub ( help('distance') ), but my suggestion would be to do clustering on the ctmm area estimates or area & tau[position] estimates (though those two are usually so correlated that it's probably better to drop one). When the animal is in a resident state, the area estimate is a genuine home-range area and is a smaller number that is closer to the data, while when the animal is in a non-resident state, the area estimate blows up because it's extrapolating whatever the current behavior is to continue forever. The advantage of the code that I am working on over vanilla clustering is that the distributions are appropriate for (0,Inf) area variables and the uncertainty is propagated through the calculation---it's kind of a clustering/meta-analysis hybrid method.

Best,

Chris

To unsubscribe from this group and stop receiving emails from it, send an email to ctmm-user+unsubscribe@googlegroups.com.

[Genevieve Finerty]

That sounds awesome -- will be exciting to see that come out! I would imagine this is a challenge that will come up again, so will be keeping an eye out for sure.

Thank you for working on the distance functions already and I think the suggestion to use the ctmm area and/or tau[position] estimates also sounds like a solid option -- going to have a play around with all of that this afternoon!

All the best,

Gen

To unsubscribe from this group and stop receiving emails from it, send an email to ctmm-user+...@googlegroups.com.