weighted or unweighted AKDEs for iRSF

[Francisco Castellanos]

Hello Chris and ctmm team,

I am analyzing data of some foxes living close to human-populated areas, and am interested in understanding when and how often they are selecting or avoiding these places.  

I have been running loops to fit ctmm, estimate monthly weighted AKDEs per individual, and run rsf.select. Once this process is done, I estimate the individual mean of the iRSFs, and plan to later fit the mean for the population. I am using this approach since fitting iRSFs seasonally or for the whole tracking period, usually results in no covariate being selected by rsf.select, which is interesting and makes sense since these are opportunistic animals that move many km per night across their home range, so I decided to capture a finer-scale behavior instead. 

The issue I am having when estimating the mean per individual is that the CIs of one covariate (anthropic_cov, a logical raster) are enormous, and some values in the summary are 0 or Inf. See the output below:

This covariate is selected and significant in some months, non-significant in others, or not selected by rsf.select. When I repeat the same process but using unweighted AKDEs, then the CIs of anthropic_cov, and the summary output is much more reasonable: 

 It is also interesting that the betas of all other covariates are very similar to the previous output. I guess this is better? Should I use unweighted or weighted AKDEs according to each animal? This is the second out of 5 foxes that is causing this problem, reason for which I decided to reach out.

I appreciate the help! 

Francisco

[Christen Fleming]

Hi Francisco,

Here are two ideas:

1. I would compare the outputs of rsf.select() rather than rsf.fit(). When you have unsupported parameters, they can cause numerical errors in the parameter estimate covariance calculation, which could be the culprit here. This is general to anything using maximum likelihood (LM, GLM, etc.).
2. If you are taking data only during the day|night when they are active, then the straightforward application of weighting will put too much probability mass on the times before and after inactivity. If weights are needed with these kinds of data, you have to contract the inactive time gaps first.

Best,

Chris