rsf.fit vs rsf.select

[Laura Franklin]

Hi community,

I'm hoping to get some clarity here, and let me preface by saying I have zero experience with RSFs - never even learned about them in classes.

I have a group of white-tailed deer that I'm running RSFs for by biological season (fawning, rearing, rut, etc.). 

I've been able to identify all habitat types present in their home ranges (10 total). I'm seeking some fine-resolution answers, which is why I haven't reclassified into fewer habitat types, but I don't love this number and will reclassify if recommended. I think I saw on here somewhere that anything more than 4 (or 6?) is pushing it.

After gleaning what I could from this user group and other resources, I've tried running rsf.select on a subset of deer individually. I've then fed the top model's formula into rsf.fit (again, for each individual). 

I've done rsf.select because it sounds like that's what's best for multiple habitat types. 

Am I on the right track? For some reason I have a weird feeling about what I'm doing. Any advice would be appreciated!

Sincerely,

Laura (PhD Candidate)

[Jesse Alston]

Hi Laura,

We advise always using rsf.select() because we have found that poorly estimated coefficients can cause problems when estimating population-level averages. This often happens, for example, when a specific cover type is not in an animal's home range, or occurs very rarely. No need for a secondary rsf.fit() command.

When using cover type data, remember that the coefficients are in comparison to a reference category (i.e., whatever cover type(s) is/are left out). A positive number is not selection for that coefficient--it is selection compared to the baseline reference category. I generally advise that folks make sure their reference category is relatively large (i.e., contains > 25% or so of an animal's locations, which ensures that selection for the reference category is well-estimated) and is visited roughly in proportion to availability (i.e., the proportion of locations in that cover type is roughly equal to the proportion that occurs in the home range or study area, which makes selection coefficients a little easier to interpret). I usually include very rare cover types (<2% of the landscape) in the reference category unless there is a priori reason to think they are biologically important, if exploratory analysis indicates that the animals really seem to seek them out, or if I am working with very large numbers of locations per animal that ensure that I have enough statistical power to detect biologically meaningful patterns of selection.

Jesse

--
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 visit https://groups.google.com/d/msgid/ctmm-user/f43b8ba8-f150-4665-8f14-f31efd4aef6en%40googlegroups.com.

--

Jesse Alston

+1 252 532 5236

jmalston.com

[Laura Franklin]

Hi Jesse,

Thank you so much for your swift response!

I have indeed struggled with my understanding of the correct usage of cover types. For the sake of processing speed, I did end up reclassifying some habitats so that now I have 3 categories (1 raster for each category). Each raster contains 1s for the habitat of choice and 0 for everything else. And, all three rasters were created from the same original so that extent, resolution, etc. are identical. I don't think they are colinear - there are 0s that account for other habitat types that I chose not to include in any category. But, I guess I'm still lost because I don't have a reference category (unless it was auto referencing by alphabetical order?). As a reference category, you suggested a rare cover type that may not be biologically meaningful, but also rather large? I don't think I interpreted that correctly. 

My categories right now are mixed hardwoods, coniferous, and mesic flatwoods. Mesic flatwoods are pretty evenly distributed and captured in the home ranges, and I don't have reason to believe they are especially meaningful. They are a nice middle ground, so to speak. Could I use mesic flatwoods as the reference group? Or, would the cover types I didn't categorize be better as the reference group? 

I could scrap all this and simply evaluate wet habitats versus dry habitats, since I'm hypothesizing that females avoid swamps during fawning but select for them during rut. If I went with this route, what would that mean for my reference category?

A couple of other random questions - I had one individual that bucked everything but the null model (next model was like 67 away). What does that mean? 

Lastly, I had yet another individual who's betas were all negative. I think I did something wrong here, but I saw somewhere that all negatives could be indicative of an animal really hunkering down. This was during fawning season and I wondered if that could be the reason. But again, I think I was off base somehow.

Thanks again!

Laura

[Christen Fleming]

Hi Laura,

If everywhere is in one of the 3 categories, then they are colinear as CAT1+CAT2+CAT3=1 everywhere.

If the selected model has no selection coefficients, then either selection was weak or there wasn't enough data to resolve it. You can see in the confidence intervals of the second model - are they so wide that they contain substantial values or not.

Negative betas on a categorical means that the reference category is being positively selected and these other categories are being negatively selected. They are relative to each other.

Best,

Chris

[Laura Franklin]

Hi Chris,

Thanks so much for circling back to this! I had put it on pause for a while but am ready to tackle it again.

It's possible that I'm misunderstanding colinearity here, but I don't think my categories are colinear. Each raster has 1s for the category of interest and 0s for the other categories AND 0s for categories that I chose to omit completely from all rasters. Can you please clarify?

I'm still confused about the reference category. As I remember, if categories are colinear, I need to identify a reference category in the arguments. Per Jesse's response, if I'm understanding correctly, the reference category should be habitat that contains > 25% of the animal's locations BUT also needs to be very rare? I don't understand how that is possible in many landscapes. 

I'm really not understanding what I should be thinking about when selecting a reference category. Could you please reiterate for me? If my categories are not colinear, how does that factor in? Can I still be intentional about what I'm using as the reference category?

Lastly, it's hard for me to fathom situations where selection was weak or there wasn't enough data. I'm using 3-4 years' worth of seasonal data (so altogether, roughly 12 months per individual) at fairly fine resolution fix rates (or so I thought; a mixture of 1, 4, and 6 for each individual, which I've accounted for in the akde process). Although, I am finding that where I thought the animals would change movement behavior during hunting season, they are actually just becoming fully nocturnal. 

I would appreciate any more assistance that you're able to offer!

Thanks,

Laura

To view this discussion visit https://groups.google.com/d/msgid/ctmm-user/24d7d7cb-193b-45ff-b0b2-0ace1026f358n%40googlegroups.com.