Number of Covariates Colext Model

[Rory Macklin]

Hi all,

I'm building colext models and am trying to find an appropriate rule of thumb to determine how many covariates I include in each submodel. I've been working off of the one-and-ten rule (one covariate for every ten "events"), but am not sure how this would be applied when multiple submodels (initial occupancy, colonization, extinction, detection) exist in the colext framework.

I see two possible interpretations:

1. You should only include 1 covariate across all submodels for every ten "events".

2. You can include 1 covariate in each submodel for every ten "events".

Also, the interpretation of "event" in this context is a little confusing to me. Would "events" just be simple captures? Or for an initial occupancy model would an "event" be a site being initially occupied, for a colonization model would an "event" be a colonization event (i.e. a site going from unoccupied to occupied etc. I'm not exactly sure how to best apply this rule.

Are there other rules of thumb that may better be applied in a colext context?

Thanks,

Rory

[Marc Kery]

Dear Rory,

I applaud your care to avoid overparameterizing a dynamic occupancy model. All too often one sees analyses where folks have, say, 20 sites and then want to add 10 or even more covariates into the model. It is good to keep in mind such rules of thumb of regression modeling as the 1 for 10 rule. At the same time, we must also remember that the challenge for these hierarchical models is even bigger than for non-hierarchical regression models, because many of these events are latent rather than observed. Thus, perhaps you should go for a 1 for >10 rule with an occupancy model.

Having said this, a far less principled but pragmatic solution is to simply try things out. That is, add covariates one by one and check out when numerical 'ugliness' shows up in the form of MLEs and especially SEs that are very large (e.g., >5), or SEs that are NaN or NA. When this happens, then you know that you have gone too far and should simplify your model.

The pragmatic approach 'happens' somewhat automatically when one follows the most important modeling rule: Start (very) simple and add complexity incrementally until you reach the desired model. To avoid p-hacking, you could a priori make a wish list of the covariates that you'd like to put on a parameter and then add them in the order of that list.

Best regards --- Marc

---

From: unma...@googlegroups.com <unma...@googlegroups.com> on behalf of Rory Macklin <rdjma...@gmail.com>
Sent: Thursday, June 26, 2025 18:23
To: unmarked <unma...@googlegroups.com>
Subject: [unmarked] Number of Covariates Colext Model

 

--
*** 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/ec8ea468-4005-48b4-bd45-0779d24f77c6n%40googlegroups.com.

[Jim Baldwin]

Just for historical context: in the ancient olden days prior to the religious rise of AIC, one occasionally generated a few random predictors in a regression and when those showed up as significant, you knew it was time to stop.

Jim

To view this discussion visit https://groups.google.com/d/msgid/unmarked/ZR0P278MB08694BB98F1D718B60A2F444EB7AA%40ZR0P278MB0869.CHEP278.PROD.OUTLOOK.COM.

[Rory Macklin]

Hi Marc,

Thanks for the response. Much appreciated insight.

If I was to apply a say 1 for 20 rule, would that mean 1 parameter in each submodel for every 20 observations, or 1 parameter across submodels for every 20 observations?

Also, should some sort of AIC-based model selection procedure weed out overparameterized models? Presumably these would have high AIC.

Best,

Rory

[Marc Kery]

Dear Rory,

my instinct would tell me the former: 1 for 20 in each submodel. The reason for this is that each parameter is estimated from different subsets of the data: psi1 just from the first year of data, then extinction between year t and t+1 is estimated by taking all sites occupied at t and seeing how many are no longer occupied at t+1. Likewise, for colonization you take sites not occupied at t and then see how many have become occupied at t+1. At least when ignoring imperfect detection, fitting the dynocc with colext should give you identical estimates as when you separate out subsets of the data and fit simple logistic regressions in each case.

But as so often, if you have a very deep interest in questions like these and some time on your hand, you might investigate them in the experimental setting of simulations.

Best regards  --- Marc

---

From: unma...@googlegroups.com <unma...@googlegroups.com> on behalf of Rory Macklin <rdjma...@gmail.com>

Sent: Thursday, June 26, 2025 22:49
To: unmarked <unma...@googlegroups.com>
Subject: Re: [unmarked] Number of Covariates Colext Model

 

To view this discussion visit https://groups.google.com/d/msgid/unmarked/270f70f8-5ab3-485f-a054-ea02ccf8f530n%40googlegroups.com.

[Marc Kery]

Dear Rory,

one more. The sample sizes are very different for the parameters in the dynocc model. Since psi1 is estimated only from the latent z states in the first year, sample size is equal to the number of sites. Next, eps and gamma are estimated from all the transitions, so sample sizes are the number of occupied and unoccupied sites, respectively, for years 1 to T-1 (where T is the total number of years). Finally, detection is estimated from every single datum, so sample size for it is the product of nsites, nyears and nreps.

Best regards  —Marc

---

From: unma...@googlegroups.com <unma...@googlegroups.com> on behalf of Rory Macklin <rdjma...@gmail.com>

Sent: Thursday, June 26, 2025 22:49
To: unmarked <unma...@googlegroups.com>

Subject: Re: [unmarked] Number of Covariates Colext Model

 

To view this discussion visit https://groups.google.com/d/msgid/unmarked/270f70f8-5ab3-485f-a054-ea02ccf8f530n%40googlegroups.com.

[Rory Macklin]

Hey Marc,

Thanks for the detailed explanation. This all makes a lot of sense - much appreciated.

Best,

Rory