Predicted habitat
The EXP approach selected optimal models with generally smaller
predicted habitat among all the approaches given we avoided over
prediction while selecting ecologically plausible models. Among the four
sequential approaches AUCDIFF approaches generally
predicted a larger area as suitable habitat for both the taxonomic
groups compared to the two ORTEST approaches when the
balance threshold was used. For some species the habitat predicted was
several times bigger than that predicted by the EXP approach, thus over
predicting. When the less restricted thresholds, considered robust for
the species with small occurrence records (Pearson et al. 2007, Warren
and Seifert 2011), are used to develop binary suitable habitat maps, as
we did here, use of ORTEST approaches may be a better
choice over AUCDIFF approaches for optimal model
selection. While the use of AUCDIFF approaches may not
be useful when a restrictive threshold, like the percentile threshold
(Galante et al. 2018), is used, since optimal models selected through
AUCDIFF approaches under predicted for some species
(Figure 5). Further studies may be required to confirm if there is a
true relationship between the thresholds used to derive binary suitable
maps and the corresponding model selection approaches used to select the
optimal models.
Though the expert approach selected more complex models judged using RM
values, ORs, AUCDIFF and average number of parameters,
the use of a less restrictive threshold value for producing binary
suitable habitat maps helped select ecologically plausible models
(Pearson et al. 2007). We suggest the use of ORTESTapproaches over AUCDIFF approaches as either the first
line of optimal model screening or by their own to select the final
optimal models since optimal models chosen by these approaches had
higher correlation with the optimal models chosen by the EXP approach
for almost all the parameters we tested. Further, though
AUCDIFF approaches selected less overfit models over
ORTEST and EXP approaches, AUCDIFFapproaches over predicted the habitat area when the balance threshold (a
less restrictive threshold) was used in our case. However, we feel there
is a need for further studies using other thresholding rules available
in the Maxent result as well as including a broader range of taxonomic
groups to assess the generality of our findings.