Maxent and presence-absence data

[bouderb...@gmail.com]

Hello, 

I read in the literature that maxent algorithm gives it best results when is applied only the presence data and not presence-absence. I want to understand the technical explanation of this point. 

Thank you in advance 

[Adam B. Smith]

Maxent does a neat trick--it first estimates the probability of the environment at presence sites by comparing their frequency to the frequency of the environmental values across the landscape (the background sites).  It then uses Bayes' theorem to invert this to get the probability of occurrence given the environment (or an index thereof--you can't actually estimate the probability of occurrence without absence data).  You can see in that the first step above that you need an estimate of the frequency of each variable in the background.  So if you were to use absence data, it would not give you an unbiased estimate of the frequency of each variable across the entire landscape.

In some cases we used biased background sites to cancel out bias in the occurrences, but otherwise the background data should still be "background" (not absences). To see why, consider the case where a species occupies all habitats of a special type, and no habitats outside that.  In this case, if we used absences, which are associated only with environments the species cannot live in, Maxent cannot use the environmental values associated with the absences to calculate the probability of the environment given occurrence because there are no such environments in the "background" (absence) data.  In practice, Maxent will still work if you use absences, but the output is not mathematically robust.

A more intuitive way to think about what Maxent does is considering he question, "I just saw a polar bear.  What kind of climate am I currently in?" Unless I'm in a zoo, I'm most certainly in a very cold climate.  You figured this out by knowing 1) that polar bears prefer cold; and 2) there are cold places on Earth (the background available to polar bears).  So now you know the "probability" I'm in a cold environment given that I saw a polar bear. But then, you invert the question (the Bayes part): "I'm in a very cold place--could I see a polar bear?"  Obviously, there are conditions (you could be at the top of a temperature/tropical mountain or in the Antarctic), but that's where the study design comes in (e.g., selecting only accessible areas from which to draw background points).

I hope that helps,

Adam

~~~~~~~~~~~~~~~~~~

Adam B. Smith, Ph.D.

Global Change Conservation Lab, Missouri Botanical Garden

Best practices in species distribution modeling: A free workshop

Biodiversity data cleaning: A free webinar

enmSdm: An R package for better, faster modeling of niches and distributions

[Mark Lawler]

Adam, while I did not ask this question, I appreciate your explanation.  I am at the beginning stages of my doctoral research which focuses on woodland caribou (Rangifer tarandus) in eastern North America and am still wrapping my head around the construction of correlative SDMs using Maxent as well as using mechanistic models (Nichemapr).

Best,

Mark

--
You received this message because you are subscribed to the Google Groups "Maxent" group.
To unsubscribe from this group and stop receiving emails from it, send an email to maxent+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/maxent/356ef702-626e-4d0b-8637-0907ff676d54n%40googlegroups.com.

[Adam B. Smith]

Glad it helped!  Honestly, my advisor (who wrote a book on Maxent and had been a theoretical physicist), read the very first Maxent SDM paper by Phillips et al. and could not recreate their modeling framework.  It wasn't clear in that paper, but at that time Maxent was proprietary because the resarch was done at AT&T labs.  Since then it's become open-source, but I have found none of the various papers that attempt to explain it very clear.

Good luck!

Adam

~~~~~~~~~~~~~~~~~~

Adam B. Smith, Ph.D.

Global Change Conservation Lab, Missouri Botanical Garden

Best practices in species distribution modeling: A free workshop

Biodiversity data cleaning: A free webinar

enmSdm: An R package for better, faster modeling of niches and distributions