Single-Season Single Species Occ. Models Fail to Converge unless Standardized

[Matthew Broadway]

Hello Group.

I apologize if this issue is redundant to previous questions. But, there are indeed over 100+ conversations to sort through to determine that.

I have been using Program MARK to fit single-season single species occupancy models for a bird species (Eastern Whip-poor-will) observed during nocturnal aural point-count surveys. We surveyed 584 sites in 2 field seasons (most of them in both seasons). Survey sites were determined using a stratified random design. Temporal replication ranges between 1-10 visits/site. We collected several Observation-level covariates to explain the detection process that included 'nuisance' variables (e.g., wind speed), and environmental variables hypothesized to influence calling behavior, and therefore detectability (e.g., moon illumination). Total individual whip-poor-will detections is ~185 and nearly identical among both years. However, sites where multiple individuals were detected are collapsed to a 0/1 encounter history (EH; MacKenzie et al. 2002).

I started by building a global detection model before reducing that to single-covariate main effects models for each observation-level detection covariate. In all cases, I standardized (z-transformed) predictor covariates and held Psi constant. Initially, all models converged and provided reasonable estimates of p and psi. However, for ease of interpretation, I wanted to run models without standardizing predictor covariates. After doing so, all models failed to converge and did not provide parameter estimates beyond the first 3-4 detection estimates (out of 10), which I thought was odd.

These issues do not arise when repeating the exact process for another species we concurrently surveyed, but had nearly 6 times the detections for (i.e., Chuck-will's-widow). In other words, it seems the parameters we use for detection are informative for a very similar species (which we expected). Hence, my thoughts are that the issues occur in the EH (underlying 0/1) data, rather than the individual covariates, as a result of closure violation and/or too few EHs where detections are only 'sometimes' made, as apposed to always/usually detected and never detected.

Unless the answer is as simple as, "Just run the models using standardization and then back-transform the model estimates for interpretation," I'm hoping someone here can provide direction on how best to investigate and diagnose the issue. I already have been checking individual covariate distributions and EH patterns. From there, I can determine how best to address it during model-fitting.

Thanks in advance and please let me know if more information would be helpful.

Cheers,

Matt

[Ken Kellner]

If you're using program MARK you should ask on the MARK forums

http://www.phidot.org/forum/viewforum.php?f=13

If you are actually using unmarked, the most likely solution is what you suggest: keep the covariates standardized. Covariates with large absolute magnitudes can cause convergence issues or other optimization problems, especially if the observations are sparse and the covariate values are far from 0. I imagine this is also true with MARK.

You can use the scale() function directly inside your formulas in unmarked which makes the back-transformation automatic when you use predict() later. See for example this thread

https://groups.google.com/g/unmarked/c/ejQqp7Qr__s/m/Rbpi9I8vAgAJ

Ken

> --
> *** 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 on the web visit https://groups.google.com/d/msgid/unmarked/821bb6b0-9aa0-4a44-904d-7172d74d80c6n%40googlegroups.com.