Re: spatio-temporal random effects vs colonisation & extinction probabilities

[Charles Emogor]

 

 

 

 

From: C.A. Emogor <ca...@cam.ac.uk>
Date: Monday, 18 August 2025 at 17:36
To: spOccupancy and spAbundance users <spocc-spa...@googlegroups.com>
Subject: spatio-temporal random effects vs colonisation & extinction probabilities

Dear Jeff, Marc, all

 

I would appreciate your thoughts on the questions below. For context, I am using spOccupancy models to understand poaching distribution across Africa (using data collected via SMART by rangers during routine patrols). As I have data across 5-10 years per site, I am using year as my primary period and unique patrols within each year as my ‘replicate’ (or secondary period).

 

1. It looks like the package does not allow for direct estimation of colonisation/extinction probabilities, but does this via modelling spatio-temporal occurrence patterns using independent spatial and temporal random effects (see Brief overview of spatial-temporal occupancy models section; https://doserlab.com/files/spoccupancy-web/articles/spacetimemodelshtml). As I understand it, this means that I cannot answer certain questions. For example, spatio-temporal RE would be appropriate to answer this research question: Where and when is poaching most likely, given landscape features and patrol effort? whereas I need to model C/E probabilities directly to answer this RQ, which I am more interested in: Do ranger patrols cause hunters to abandon sites (extinction) or shift to new areas (colonisation)? Is there a way to get at the latter in spOccupancy?

 

2. My current model would not converge. Although the detection part seems reasonable (Rhat of 1 and ESS >500), the occupancy part is bad (Rhat of 1 to 20 across covariates and ESS of max 250; see attachment). I started with weakly informative priors and have tried a) setting the priors based on the raw occupancy and b) increasing the number of simulations. Are there any other practical steps? Do you reckon the approximation is good? Is the model guaranteed to converge if I run it for a very long time?

 

3. My current model is a multi-season single-species model. I want to extend it to a multi-season, multi-species model so I can also account for species distribution. Ideally, I want to capture any correlations between species distribution and the occurrence of poaching. I'm thus unsure how best to implement the model: should I extract species occurrence (either presence/absence or count) per grid and use it as a predictor, or include species presence/absence as a response variable within a joint modelling framework?

 

I would appreciate any thoughts you have.

 

Regards

Charles

.....

Charles Emogor, PhD

Schmidt Science Fellow

Department of Computer Science and Technology, University of Cambridge &

John A. Paulson School of Engineering and Applied Sciences, Harvard University

https://pangolino.org/ | https://charlesemogor.com/

 

 

[Jeffrey Doser]

Hi Charles,

Apologies for the delay. Here are some quick thoughts on your three points:

1. Unfortunately, that sort of question is not really possible with the existing spOccupancy functionality. While you could derive post-hoc estimates of colonization and extinction, you wouldn't really be able to relate that to any covariates without then needing to run some other analysis, which would not be ideal. If that is your main interest, your best bet is probably to try and use the colext function in unmarked, or if you need a multi-species model, you may have to implement a multi-species dynamic occupancy model in JAGS and/or nimble (I'm sure you could find some code from a paper somewhere that does that).
   
2. It seems like there is certainly some sort of problem with the model that is preventing convergence. There is no guarantee that convergence will happen, particularly if the model is overly complex or there is some other reason for not being able to identify all the model parameters. I would suggest you start the model to be much simpler. See if you can first just get an intercept only model to work, then gradually increase model complexity to see what is causing the problem. A couple of things I noted is that the "accessibility" covariate in particular is showing signs of not being able to be estimated, so you may check that to see if there is a reason why that might be the case (e.g., not enough variability in the values). The two random effect variances are also extremely large, which indicates some sort of identifiability problem trying to estimate those.
3. This probably depends on how many species you are working with. Is it just one species that you want to look at its relationship to poaching? If so, it might be easier to just use presence/absence as a covariate, with the caveat that that covariate would not account for the fact that there may be imperfect detection in that species distribution. If you have a lot of species, modeling jointly would probably be the way to go (could use stMsPGOcc to do that).
   

Jeff