occuRN Model Averaging Inquiry

40 views
Skip to first unread message

Caroline Pollan

unread,
Aug 4, 2026, 7:37:23 AM (3 days ago) Aug 4
to unmarked
Hi everyone!
I'm currently working on a bat bioacoustics project, and have determined that a Royle-Nichols model is the best approach for analysis. As predictor variables, I have two detection covariates and three site covariates. I have fit models with all combinations of the aforementioned variables (none of these variables are correlated according to a VIF test), including interactions, as these combinations are all entirely possible given my study system. When ranking models by AIC value, there are ten models that have a delta AIC value of less than two, leading me to employ model averaging. My question is: how should I go about dealing with interactions in model averaging for occuRN? Should I leave models containing interactions out of averaging, and consider them separately?

Another thing I should note is that, in this case, one of the models included in the top ten was composed of only detection covariates and no site covariates. Furthermore, inclusion of any combination of site covariates with no detection covariates did not achieve a lower AIC than the true null model. From here, my conclusion would be that these site covariates are not informative, but am not entirely certain if that is a conclusion that can be made from that information.

Any advice would be welcome! 
Thank you,

Caroline

Marc Kery

unread,
Aug 4, 2026, 7:47:56 AM (3 days ago) Aug 4
to unma...@googlegroups.com
Dear Caroline,

just out of curiosity, can you share a summary of the unmarked data frame ?

Thanks and best regards  -- Marc


From: unma...@googlegroups.com <unma...@googlegroups.com> on behalf of Caroline Pollan <ca...@rice.edu>
Sent: Tuesday, August 4, 2026 04:47
To: unmarked <unma...@googlegroups.com>
Subject: [unmarked] occuRN Model Averaging Inquiry
 
--
*** Three hierarchical modeling email lists ***
(1) unmarked (this list): for questions specific to the R package unmarked
(2) SCR: for design and Bayesian or non-bayesian analysis of spatial capture-recapture
(3) HMecology: for everything else, especially material covered in the books by Royle & Dorazio (2008), Kéry & Schaub (2012), Kéry & Royle (2016, 2021) and Schaub & Kéry (2022)
---
You received this message because you are subscribed to the Google Groups "unmarked" group.
To unsubscribe from this group and stop receiving emails from it, send an email to unmarked+u...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/unmarked/6a46779d-5a9f-4638-815d-b1b3051a550an%40googlegroups.com.

Caroline Pollan

unread,
Aug 5, 2026, 2:00:51 PM (2 days ago) Aug 5
to unmarked
Yes, of course!
The detection covariates I am using are "time" and "clutter", and the site covariates are "Elevation_m", "Forest", and "Canopy_avg". I am scaling the numeric covariates (elevation (m), canopy average (%), and clutter (m)),when including them in models.

Screenshot 2026-08-05 at 1.53.25 PM.png
Thank you so much,
Caroline

Marc Kery

unread,
Aug 6, 2026, 3:06:14 AM (24 hours ago) Aug 6
to unmarked
Dear Caroline,

thanks, but that unfortunately shows you don't have enough data to do much covariate modeling.

Occupancy, RN, Nmix models etc can be viewed as the combination of two regression models: one for the true state (presence/absence, abundance: state model) and another for detection (observation model). These regressions can consist of an intercept only, i.e., when you don't add any covariates.
 
State and observation models have different sample sizes: whereas for the state model it is the number of sites, the relevant sample size for estimating parameters in the observation model is (normally) the total number of visits over all sites. In the regression modeling literature people have come up with rules of thumb about the sample sizes needed to estimate one parameter, and I believe typical values offered are between 5 and 20.
 
And this is for simple regressions, where the data points are directly observed. In the state model of occupancy or Nmix models, the values of the states are *not* directly observed but must be estimated. Hence, to estimate a single parameter in the state model, we may need perhaps 10 to 40 sites.
 
As a consequence, with 12 sites you should not consider any covariates in the state process but have an intercept-only model. For the observation model the situation is a little less bleak and perhaps you could consider 1-2 covariates for it.
 
Sorry to be the bearer of bad news ...

Best regards  -- Marc
 

From: unma...@googlegroups.com <unma...@googlegroups.com> on behalf of Caroline Pollan <ca...@rice.edu>
Sent: Wednesday, August 5, 2026 20:00
To: unmarked <unma...@googlegroups.com>
Subject: Re: [unmarked] occuRN Model Averaging Inquiry
 
Reply all
Reply to author
Forward
0 new messages