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

Abstract

The widely used “Maxent” software for modeling species distributions from presence‐only data (Phillips et al., Ecological Modelling, 190, 2006, 231) tends to produce models with high‐predictive performance but low‐ecological interpretability, and implications of Maxent's statistical approach to variable transformation, model fitting, and model selection remain underappreciated. In particular, Maxent's approach to model selection through lasso regularization has been shown to give less parsimonious distribution models—that is, models which are more complex but not necessarily predictively better—than subset selection. In this paper, we introduce the MIAmaxent R package, which provides a statistical approach to modeling species distributions similar to Maxent's, but with subset selection instead of lasso regularization. The simpler models typically produced by subset selection are ecologically more interpretable, and making distribution models more grounded in ecological theory is a fundamental motivation for using MIAmaxent. To that end, the package executes variable transformation based on expected occurrence–environment relationships and contains tools for exploring data and interrogating models in light of knowledge of the modeled system. Additionally, MIAmaxent implements two different kinds of model fitting: maximum entropy fitting for presence‐only data and logistic regression (GLM) for presence–absence data. Unlike Maxent, MIAmaxent decouples variable transformation, model fitting, and model selection, which facilitates methodological comparisons and gives the modeler greater flexibility when choosing a statistical approach to a given distribution modeling problem.

Comparison of Habitat Suitability Models for Haemaphysalis longicornis Neumann in North America to Determine Its Potential Geographic Range

“…Figure S1: Frequency of Observed Presence (FOP) plots for the predictors of interest (WorldClim bio1-annual mean temperature; bio5-max temperature of warmest month; bio12-annual precipitation; bio10-mean temperature of warmest month; and bio11-mean temperature of coldest month (temperature in • C*10 and precipitation in mm); GEZ-global ecozone; ENVIREM continentality (in • C) for the East Asia region with 325 presences and 980 background points. These plots were created via the R package MIAmaxent [38], where the dots are the values of the predictors at the given locations, the red line a smoother regression line, and the background distribution approximate the data density; Table S1: Results of estimating the best MaxEnt model features and regularization (rm) for each geographic area of interest using the ENMeval R package [39]. Results are based on the random 5-fold method for data partitioning, where background points were randomly selected from the area of unsuitable habitat modeled from the BIOCLIM algorithm (a classic presence-only climate envelope model), and settings that primarily minimize AICc (i.e., ∆AIcc = 0) were selected for our best models (in bold); however, the AUC metrics and OR (threshold-based omission rates for test localities) metrics were calculated to select less complex models (when compared to Frequency of Observed Presence plots) and lowest number of parameters if those were giving similar or higher AUC and lowest OR because a low OR indicates less overfitting.…”

“…Results are based on the random 5-fold method for data partitioning, where background points were randomly selected from the area of unsuitable habitat modeled from the BIOCLIM algorithm (a classic presence-only climate envelope model), and settings that primarily minimize AICc (i.e., ∆AIcc = 0) were selected for our best models (in bold); however, the AUC metrics and OR (threshold-based omission rates for test localities) metrics were calculated to select less complex models (when compared to Frequency of Observed Presence plots) and lowest number of parameters if those were giving similar or higher AUC and lowest OR because a low OR indicates less overfitting. [L: linear, Q: Quadratics, H: hinge]; Table S2: Results of the nested MaxEnt-type models built during the forward DV selection using the MIAmaxent R package [38], where DV is the derived variables from the original ones using a specified transformation [Linear, Quadratics, Monotonous, Forward or Reverse Hinge, or Threshold for Continuous variable and binary for Categorical variable] that balance complexity of model with its fitness. Alpha = 0.005 was used to set the threshold for the amount of variation a DV must explain to be kept, i.e., P < alpha.…”

“…As the occurrence records were of presence-only data, maximum entropy distribution modeling, or MaxEnt modeling was used to create habitat suitability maps of H. longicornis in North America. Statistical modeling was done by running MaxEnt (New York, U.S.A.) [31] in R [34] within Dismo [37], MIAmaxent [38], and ENMeval [39] packages. The raster stack containing the environmental predictors was separated into four geographic areas of interest: (1) the current range of H. longicornis in both East Asia and Oceania (entire distribution), (2) the native range of H. longicornis in East Asia (East Asia), (3) the introduced range of H. longicornis in Oceania (Oceania), and (4) the parthenogenetic range of H. longicornis in East Asia (native parthenogenetic).…”

“…The environmental data for all four areas of interest were compared using box plots. The frequency of observed presence (FOP) plots [38] for the predictors of interest were also analyzed to determine whether the patterns of occurrence specific to the study area were compatible with the ecological knowledge of the H. longicornis ( Figure S1).…”

“…For the second modeling step, we used 12 bioclimatic variables from WorldClim Version 2 [32] and 3 variables (i.e., annualPET, Thornthwaite aridity index, and continentality) from ENVIREM [33]. These variables were used to find the best subsets of predictors for MaxEnt modeling using a forward stepwise selection process [38]. Using the subsets of predictors identified, a series of candidate MaxEnt models (Table S2) were developed using the same methodology described for the first modeling step.…”