The MIAmaxent R package: Variable transformation and model selection for species distribution models

Area of Science:

• Ecology
• Computational Biology
• Biostatistics

Background:

• Species distribution models (SDMs) are crucial for understanding biodiversity patterns.
• Maxent software, widely used for SDMs with presence-only data, often yields high predictive accuracy but lacks ecological interpretability.
• Maxent's reliance on lasso regularization for model selection can result in overly complex, less parsimonious models.

Purpose of the Study:

• Introduce the MIAmaxent R package, an alternative to Maxent for species distribution modeling.
• Provide a statistically robust approach that prioritizes ecological interpretability through subset selection.
• Facilitate a deeper integration of ecological theory into species distribution modeling.

Main Methods:

• Developed the MIAmaxent R package, employing subset selection instead of lasso regularization for model selection.
• Implemented variable transformation based on expected occurrence-environment relationships.
• Included tools for data exploration and model interrogation within an ecological context.
• Integrated two model fitting methods: maximum entropy for presence-only data and logistic regression (GLM) for presence-absence data.
• Decoupled variable transformation, model fitting, and model selection for enhanced flexibility.

Main Results:

• MIAmaxent typically produces more parsimonious and ecologically interpretable species distribution models compared to Maxent.
• The package's decoupled approach allows for greater flexibility in statistical methodology selection.
• Variable transformation and model interrogation tools support a more theory-driven modeling process.

Conclusions:

• MIAmaxent offers a valuable alternative for researchers seeking ecologically interpretable species distribution models.
• The package's design promotes a stronger link between ecological theory and statistical modeling.
• Decoupling modeling steps enhances methodological transparency and user control in species distribution modeling.