New approaches to species distribution modeling with Maxent: Rethinking model interpretation, model complexity and prior assumptions

Background/Question/Methods

Maxent is one of the most popular species distribution modeling methods, with over 400 published applications in the just the last six years. Maxent users are confronted with a wide variety of options when fitting their models, from the multiple options and settings available in the software to which input datasets to choose. However, the default settings are often chosen as a consequence of unfamiliarity with maximum entropy models, even though alternatives may often be more appropriate when connected to specific ecological questions. To explore Maxent’s assumptions, we demonstrate the variability in model output that can result from altering model settings and offer suggestions for choosing these settings.

Results/Conclusions

Maxent models are capable of predicting the relative probability of occurrence, but not the absolute probability of occurrence. Considering this, and the quality of opportunistically collected occurrence data and coarse resolution, remotely sensed environmental data, predictions must be interpreted cautiously when creating range maps and using them to answer complex ecological or evolutionary questions. In many cases, Maxent is best suited for hypothesis generation and asking better questions, not answering them.

To date, variable selection methods explored for Maxent derive almost exclusively from machine learning perspectives, which focus more on complex pattern recognition than on producing easily interpreted models. We outline a more general approach, based on constructing simpler models motivated by specific ecological questions.

In Maxent, different prior assumptions (null models) are reflected by methods of background sample selection and accounting for sampling bias. We relate these subtle cases of prior specification to the more general case, where ‘ecological’ priors can be used to incorporate different ecological assumptions or output from other models. Ecological priors are used to improve predictions of models for invasive ranges using native range data and models accounting for dispersal limitation. When ecological priors are applied to understand the spread Celastrus orbiculatus, an invasive liana, across New England (USA) over the last century, we find that greater spread is predicted in northern New England than models using default settings.