# How maxent's algorithm uses data to predicte distribution of species

Using the same list of occurrences with present and future bioclimatic variables (worldclim) I get two obviously different distributions. In the case of the future projection, I assume that the occurrences are the same as now. The same climatic conditions as now, however, will not be the same as the 2070 model, so in 2070 the species may no longer exist. How can I correct this factin the data analysed? Not only will bioclimatic variables change but land cover and other variables will also not be the same. Can I consider the second result (biocliamic variables of 2070) as the distribution of the species in the future?

I am preparing my master's thesis in geography at the University of Bologna. I would like to use an indicator species (e.g. bees) to check the effects of climate change in the Apuan Alps over the next 50 years.

• Renner & Warton (2013), Yackulic et al (2012) and Royle et al (2012) have all demonstrated serious issues with MaxEnt. Of particular note is the models mathematical equivalence to GLM's. This raises questions regarding its nonparametric characteristics and the quality of nonlinear relatioships fit throug a step function. Given that, in the case, you are dealing with a conditional losgistic function, the ability to project into new parameter space, representing novel climates, is not at all supported. Commented Oct 19, 2020 at 18:05

## 1 Answer

You can say it's a model or an estimate of that species might be distributed in 2070, but you don't know for certain that is where this species will be in 2070. The bioclimatic variables for 2070 are obviously a model too (so any unexpected changes in them might change species distribution), as well as there might be aspects of species distribution you haven't taken into account. For example a massive die off driven in part by nonclimatic vectors (i.e. sea star wasting disease) wouldn't be taken into account by your model. Likewise habitat lose due to landcover change may not be captured (i.e. increased development or other intentional landcover change).

If you choose to use this model to make conclusions related to species distribution in 2070, make sure you document how you came up with your model. What variables you used, where you got the 2070 versions, how your model estimates species distribution based on these variables, any other tweaking you did of your model, and any variables you think might impact species distribution that you didn't take into account. No model is perfect, or captures real life 100%, but that doesn't mean models never get used. It just means you have to document your model well, including detailing its strengths and weaknesses.

For further details about specific use of the MaxEnt technique, Merow, Smith, and Silander (2013) gives a practical guide on how to use the MaxEnt model for modeling species distribution, including some of its strengths and weaknesses.

• Thank you very much, it was the confirmation I was looking for but I don't understand how maxent calculates future occurrences based on current ones. Probably MaxEnt gives me a general projection of the future distribution. Thank you very much, it was the confirmation I was looking for but I don't understand how maxent calculates future occurrences based on current ones. Probably MaxEnt gives me a general projection of the future distribution. I have used worldclime and am building other variables (ground covers, etc.). Commented Oct 19, 2020 at 12:49
• So that sounds like your question really is about how MaxEnt works as a spatial algorithm, not more general use of modeling species distribution over time. You might want to edit your question to reflect this. I found a paper related to this which I will include in my answer. Commented Oct 19, 2020 at 13:51
• Yes! I read it (Merow, Smith, and Silander (2013)). I will modify the question as you suggested. Thanks a lot. Commented Oct 19, 2020 at 13:59
• I would point out that MaxEnt is NOT a spatial algorithm. Just because a model can be spatially estimated does not make it spatial. There is nothing in the equations that incorporate spatial structure or die-off into the estimates. You could approximate a naive spatial model by including spatial coordinates or a parameter representing some type of spatial structure but, this is occurring in the parameter space. You would be much better off just specifying a GLM and then using a mixed effects model or GAM to account for spatial effects with a Monte Carlo to model future climates. Commented Oct 19, 2020 at 18:12