Spatial capture-recapture in Stan

Hi all,

I’ve translated most of the spatial capture-recapture models from the 2013 book “Spatial Capture-Recapture” by Andy Royle, Richard Chandler, Beth Gardner, and Rahel Sollmann from JAGS to Stan:

Spatial capture-recapture models are used to estimate density of (usually wildlife) populations monitored by camera traps, acoustic sensors, hair snares, and a variety of other methods. They extend traditional capture-recapture methods by explicitly modeling spatial activity centers of individuals, and in some cases, explicitly modeling movement over space and time at the individual level.

This plot might help provide some intuition:

Given some detections (black dots) of an individual in a trapping array (grey crosses), we can sample from the posterior density of the location of each individual’s activity center in space (the red 2d density), and estimate overall density in a spatial region (potentially as a function of landscape-level features).

Anecdotally the Stan implementations seem to be faster than JAGS, and Stan’s sampling diagnostics reveal issues with some of the more complex models.

This type of model might also be of interest to people who work on other types of latent point process models.


@mbjoseph very cool!

@jonah, @stevebronder, this looks like something that’d make sense to link to from “Specialized Field Guides” here: . How do we do that?

1 Like

Thanks @bbales2!

I suspect it might be better to link to this repo from the existing Stan Ecology page (that is already linked to in the Stan docs). I can shoot at PR over to our Stan Ecology GitHub repo, and get it on that site.


It’s on there now! Along with the marginalizing discrete states.


Are these by any chance the 2-d partially observed t-distributed random walk models?

yes @sakrejda - that was one of them!

Edit: Kind of - IIRC both of these models raised issues:

The first assumes that instantaneous locations u are normally distributed around activity centers s:

The second assumes that instantaneous locations evolve by bivariate Gaussian random walks:

Those two are probably fixable. The t/cauchy distributed random walk models in this context have very multi-modal posteriors for the location parameters even in the 1-D case (with many of the ecological datasets I’ve seen) so they’re more of a research project than an easy translation or usable model.