Lurestick openCR

11 views
Skip to first unread message

stefano palmero

unread,
Feb 11, 2026, 5:46:23 AM (11 days ago) Feb 11
to secr
Hi all,
I have a dataset with near-perfect detection because of baited luresticks. a high survival and a very high density (1 animal per km2).
I am trying to run openCR.fit models on a long-term dataset (9 years). The CJSsecr works perfectly, but the JSSAsecrD produces incredibly inflated results for sigma and near-to-0 densities. I have been discussing this with AI and, apparently, a JSSA model does not work in my case because: 

1. Core issue: near-perfect detection
  • Your lure-stick data have very high encounter probabilities.

  • Most animals are detected in almost every session, often multiple times.

  • In JSSA, density is estimated from the number of animals that were never detected.

  • With nearly all animals observed, there is essentially no missing information for the model to estimate density.

Consequence: the likelihood cannot separate lambda0, sigma, and D, leading to:

  • lambda0 → ∞

  • sigma → ∞

  • D → 0


Does this make sense to you?

Thank you very much

Murray Efford

unread,
Feb 13, 2026, 2:37:29 AM (9 days ago) Feb 13
to secr
Not really. Your AI seems to be scrambling trad (non-spatial) and spatial concepts of detection probability ('encounter' is the giveaway - that's a trad term). It's nonsense to say density is estimated from the number of animals never detected: that number is by definition unknown. JSSA is the non-spatial variant that doesn't estimate density. As always, AI makes stuff up.

I don't know lure sticks. I would guess there is a positive learned response, and that would explain what you describe. Very high recapture rates give you essentially known-fate data, which are good for survival but uninformative for first-capture probability. However I don't get why sigma-hat should go to infinity. It should be easy to simulate what the AI claims. Also, even if the underlying probability of first detection is high (high lambda0) we expect peripheral animals to have low probability of detection (given unbounded habitat).

On the side - CJSsecr is not recommended because it's not really a spatial CJS analogue - try PLBsecrx.

Cheers
Murray

stefano palmero

unread,
Feb 13, 2026, 5:12:02 AM (9 days ago) Feb 13
to Murray Efford, secr
Hi Murray,
Thank you very much for the response. I thought so too, that the AI was making up stuff.
My goal was to compare closed and open densities, expecting the latter to fluctuate less compared to the former.
Only, the JSSAsecrD does not work on my dataset, and I do not understand why.
In the end, I estimated densities as a derived parameter by applying derived() to a JSSAsecrl model with lambda~t, and the numbers make sense.
Do you think that a JSSAsecrlCL (PLBsecrl) could provide me with all the information that I need (lambda, phi and derived density and also f)? 

Best

Dr. Stefano Palmero


Department of National Park Monitoring and Animal Management 

Freyunger Str. 2

94481 Grafenau

Germany



--
You received this message because you are subscribed to the Google Groups "secr" group.
To unsubscribe from this group and stop receiving emails from it, send an email to secrgroup+...@googlegroups.com.
To view this discussion, visit https://groups.google.com/d/msgid/secrgroup/545f6444-039f-4284-86ed-03089e6404den%40googlegroups.com.
Reply all
Reply to author
Forward
0 new messages