Correct: favourability does not estimate true prevalence, which (to my knowledge) can’t be known except through very thorough sampling (rarely possible at the scale of species distributions). But since sample prevalence is always wrong to some point, favourability has at least the advantage of removing its effect from model predictions, thus levelling them out so that they are quantitatively comparable across species and regions, regardless of (right or wrong) prevalence.

General workflow was: 1) Simulate a virtual species in a given area, resulting in some true prevalence (e.g., species covers 30% of the study area), 2) Sample presence and absence points from the true (simulated) range, varying the observed prevalence (ratio of presence to absence points selected). 3). Fit random forest models and recover an estimate prevalence (e.g., SDM suggest species covers 60% of the range – not good!). Sometimes with weighting, sometimes without weighting.

In the end, there was no clear way for the model-estimated prevalence to match the true (simulate) prevalence, unless you know the true prevalence (which in real life is never possible) and adjust the observed prevalence (ratio of select pres/abs points) accordingly. I stopped working on the project, but it remains curious in my head!

This post was interesting, my hope before reading was that the favorability function might be addressing my issue of estimating the latent true prevalence from only observed prevalence and environmental relationships. But that doesn’t seem to be the case, correct? Any thoughts on this?

Thanks!

It looks like your current species (i.e. your current value of `s`) has no data on GBIF. Is it correctly spelled? I’ve slightly altered the script in the post, to only try to make that data frame if data exists, so that the loop is not interrupted by an error. See also the note I added on the post, just before the R code. The (slightly updated) script should still work, though.