Data gaps and anisotrophic vs anisotrophic model choice

55 views
Skip to first unread message

Nell

unread,
Feb 15, 2021, 12:27:35 PM2/15/21
to ctmm R user group
Hi Chris,

Thanks so much for creating this package, and for giving up so much of your time to this group.

I’ve read in your help files and messages to other people that gaps in the data weren’t too much of a problem when working with continuous time, but I was wondering if there was any limit to this. I’m working with domestic cats, tracking each for a minimum of three weeks. While for some cats this was continuous, for others this 21 days of tracking was spread out over two or three calendar months. Due to fieldwork limitations, some of my cats were tracked for two weeks, then there was a gap of a month, then they were tracked again for another two or three weeks.  I suspect this has led to higher levels of variation in the latter sections of my variograms, and makes my default variogram.fit() a little high (I’m reluctant to manually decrease them as I’d like to keep the process as similar as possible between all my cats).

1.       Is this level of gap acceptable for ctmm?

2.       Variogram.fit sometimes only plots the initial part of the tracking (see the variogram.fit() left vs plot(variogram) right).  I can't tell if variogram.fit() is just operating on that first section, or the entire tracking period?

variogramfit.pngvariogramplot.png

3.       With the ctmm.select() function, I’ve noticed in some of your help files that you recommend anisotrophic models over their isotrophic counterparts. I was wondering whether this has a limit? At what difference in AIC should I revert to choosing isotrophic models, if at all?


Thanks so much in advance for your help,

Nell

Christen Fleming

unread,
Feb 15, 2021, 2:47:45 PM2/15/21
to ctmm R user group
Thanks Nell,

Gaps aren't an issue in the model fitting, beyond the loss of information and widening of confidence intervals. Gaps are an issue for variograms, and make them ugly or chop off early, like you've seen.

Because they are nested models, if the AIC difference is small then the covariance ellipse should be approximately circular, and both are plausible models. So I would just take the top model model as long as you have abundant data. AIC is more problematic with colinearity, small sample sizes, wildly different models, etc..

Best,
Chris

Nell

unread,
Feb 18, 2021, 6:49:25 AM2/18/21
to ctmm R user group
Hi Chris,

Thanks so much for getting back to me, I really appreciate it. That's great to hear about the gaps and model fitting.

Just to clarify, when you say gaps are a problem for variograms, is that just aesthetic or do I need to do something special to ensure variogram.fit() has functioned correctly before I feed the parameters into my AKDE models?

Regarding the AIC difference between anisotrophic/isotrophic, for a general rule of thumb would an AIC difference of 2 be regarded as just cause for choosing isotrophic over the preferred anisotrophic? I know you've said it relies on the abundance of data - I have 30 animals and >25,000 data points. Some of my anisotrophic models are ~3.5 AIC below their isotrophic counterparts.

Thank you so much for your help. It's a really great package! I've been recommended it by lots of people.

Nell

Christen Fleming

unread,
Feb 18, 2021, 4:15:44 PM2/18/21
to ctmm R user group
Hi Nell,

The problem is aesthetic, but in those cases you might also want to check the GUESS output for plausibility (I've tried to code it to be sensible regardless). You can also try to squeeze some improvement out of the variograms with its various arguments (dt, res, fast, ...) and by averaging multiple variograms together (from a similar population).

Regarding slight AIC differences with abundant data and nested models, I doubt that decision criterion really matters. I usually just take the lowest AIC, for making predictions, unless I'm in one of the problematic regimes that I mentioned before. Have you tested this choice to assess the difference in the final outputs? If I am wrong, please let me know!

Thanks,
Chris

Nell

unread,
Feb 25, 2021, 6:05:59 AM2/25/21
to ctmm R user group
Hi Chris,

Thanks for the clarification on the variograms!

Regarding the AIC, most differences are very large, but an individual where the difference is small, I've run the AKDEs on both the isotrophic and the anisotrophic. The differences in the estimates are really small (see below), as you say, so I'll just keep using the lowest AIC.

Thanks so much for your help! I have another question about IID but I'll create a new query for that so it's more searchable for others.

Thanks Chris, both the package and the support for it are amazing.

Nell



Difference in AIC between OU isotrophic (top model) and OU anisotrophic (third best model, AIC = 3.516) creates AKDE estimates of:

> summary(akde.1iso)

$DOF

     area        bandwidth

 511.9064  561.1849

 

$CI

                                                low      est     high

area (square meters) 4121.442 4503.138 4901.494

 

> summary(akde.1aniso)

$DOF

      area     bandwidth

 510.2839  561.0077

 

$CI

                                               low      est     high

area (square meters) 4116.126 4497.976 4896.521

Reply all
Reply to author
Forward
0 new messages