I have spent the entire week reading spatial capture-recapture papers. I've seen some good and some bad things and for myself as much as anything else I want to summarise my experience as a statistician reading these papers 🧵
So in no particular order... here's a bunch of stuff I think about the SCR literature. I was specifically looking at any papers trying to detect spatial variation in animal abundance and possible causes of this variation.
Many papers do SCR and then something else (e.g occupancy, resource selection functions). I often didn't understand the something else. What is the model? What is the response variable? There is a strong tendency that writing [generic name of big family of methods] is enough
There may be extremely detailed description of the SCR data but then only a few lines on the "something else" they did. Not enough to figure out what the actual methods were or even what the data they used was.
A common approach is to analyse data from different sites separately, not in a joint model, and then compare density estimates at each site with "what we know about those sites" to explain variation in density.
I am sympathetic to this, it means you perhaps don't fall prey to the narrow focus of "spatial covariates we happen to have access to" (cough landcover maps cough) and you can be quite flexible in considering causes and drawing conclusions
The downside is the key inference of the paper is completely informal. Your paper is about differences between sites but you didn't model those differences within a statistical framework.
SCR is largely untouched by methods to deal with spatial autocorrelation (this is what one chapter of my thesis is looking at - spoiler, it's hard). Often the "something else" model mentions the importance of this and does something like splines / restricted spatial regression.
So this is a clear gap. Ecologists know unexplained autocorrelation is something to consider in a GLM but they don't have methods to do it within SCR. I'm glad I've convinced myself this thesis chapter I'm writing is actually useful even if I didn't solve all the problems.
There are all sorts of reasons ecologists want to learn about spatial heterogeneity. To learn about habitat associations, responses to disturbance (historic or present day), responses to management efforts, identify key areas of concern, hotspots etc...
...but SCR was mainly designed to estimate abundance, not really any of that other stuff. It really is very robust at doing this (with a good sampling design). All the rest is tacked on and faces all the same challenges as species distribution models, except...
...sample sizes are often small. Yes you have load of recaptures but in the end there's only 16 tigers or whatever. Try doing a SDM on 16 observations. Try detecting habitat associations, disturbances etc, when you have idealised each individual to a point in a landscape but...
...locations of points are not that informative unless you have a lot of them (sorry to say this since point patterns are basically my PhD). So if you want SCR to be a wrapper around a SDM then you have to think about sample size.
On a related note: if I've only got 16 individuals, how much spatial autocorrelation is there? Unless clustering is strong you probably can't detect it and so you don't really need to model it. But you should be able to try and model it and see if you need it.
Often the "something else" analysis feels like something they could have included in the SCR model. Why model detections at camera traps as binomial GLM with spatial covariates when you can have an inhomogeneous Poisson process model within the SCR model?
My biggest problem with doing occupancy/GLM to analyse spatial variation and then doing homogeneous density SCR is you have chucked away the recapture information for no reason. Just do inhomogeneous Poisson process since you have the (expensive) data to infer abundance.
Telemetry data on the other hand can't readily be incorporated into the point process model (since it's about movement around the point). Decisions about what to use around an activity centre are not the same as the overall location of the activity centre in a landscape...
You could argue fine scale movement decisions is everything and that an activity centre is never really "selected". The activity centre is a really weird concept in my opinion. In papers with SCR and telemetry data I felt this discussion was missing...
...and in cases where the data were analysed jointly I usually wasn't sure exactly the relationship between the SCR and RSF models and which parameters were shared between them and how they were modelled. This is a big downside to not mathematically describing the model.
There are a lot of papers that want to do landscape ecology with SCR-ready data but they seem unaware of what can be done already within SCR framework. That's why they do "something else" and SCR is (for them) just for abundance.
We can go a long way to simplifying some of these analyses just by using the inhomogeneous poisson process. My work extends to log-Gaussian cox process but that brings a lot of new things to worry about.
I'm not sure this is actually what they did but I read numerous papers that seemed to use MaxEnt with binomial response of detections at traps. This seems wild to me. The locations of traps was by design, you can't view those as presences and add pseudo-absences...
... then away you go with MaxEnt and spatial covariates. That is genuinely bonkers. I saw the animal there because the animal was there AND I decided to put a trap there. If I put traps elsewhere nearby I'd probably see them there too. Can anyone defend this approach to me?
I guess I'll end with: ecologists always want to do more with a model than the intended use. This isn't a bad thing necessarily. But it does mean a lot of weird things get published. It's fertile ground for an applied statistician to help out.
Okay actually that is not the end. Some papers try to deal with heterogeneity by a priori restricting the state space to "potential habitat". I have numerous problems with this but the main one is that an activity centre doesn't have to be in a useable habitat...
Consider an extreme case of an animal living on the shores of a circular body of water. It just roams around this body of water (this animal can't swim). Activity centre is plonk in the middle of the lake.
I call this the donut space use model. The centre of gravity of a donut isn't part of the donut, it is the hole in the middle. Please don't DM me about whether the hole exists or is "part of the donut". I am not a philosopher. It's definitely not made of gooey batter though
Much more common example in the wild - animal that uses high elevation rugged terrain, perhaps end of a valley where high elevation is a horseshoe around the flat valley. Cats never use the flat valley but needs to be in the model as potential place for activity centre.
I return to this: an activity centre is a weird thing. An inhomogeneous Poisson process to describe distribution of weird things is also hard to think about.
Another example: patchy forest habitat. Animal moves quickly between patches. Activity centre could quite easily be in between, in land you considered "unsuitable"
You can follow @ASeatonSpatial.
Tip: mention @twtextapp on a Twitter thread with the keyword “unroll” to get a link to it.

Latest Threads Unrolled: