Model setting
We used Maxent v 3.4.1 (Phillips, Dudík and Schapire) to model the
species distributions. We used Maxent’s current default output format
Cloglog since it gives a better result over logistic when bias
correction is used (Phillips, 2017;
Phillips et al. 2017). We used
four sets of FCs resulting from the use of individual FCs independently
or in combination with other FCs, namely (i) linear (L), (ii)
linear-quadratic (LQ), (iii) hinge (H) and (iv) linear-quadratic-hinge
(LQH) after Galante et al. (2018) to build models given our species had
only a small number of occurrence records (Table S1). Higher FCs were
found to produce less complex (i.e., lesser number of parameters) and
less overfitting (i.e., lower omission rates) models when occurrences
were small (Radosavljevic and Anderson 2014). Specifically, the hinge
feature was better for species with small occurrence values (Galante et
al. 2018).
RMs with values less than default produce models which are overfit to
occurrence data and are not well generalized, while larger RMs would
produce spread out and less localized models (Phillips, 2017).
Though Radosavljevic and Anderson
(2014) also observed a slight peak in the model discriminatory ability
around the default RM, they found substantial reduction in overfitting
when RMs of two to four times that of the default were used. However,
they also found both the model quality and the overall discriminatory
power declined when RMs were above 4 (Radosavljevic and Anderson 2014).
Hence, we used conservatively 11 different RM values, namely .25, .5, 1,
1.5, 2, 2.5, 3, 3.5, 4, 4.5 and 5.
We used Maxent’s default replication method of cross-validation since it
is a better replicate option with small occurrence data. It randomly
splits occurrence data into folds and uses all the folds in turn to
build and evaluate models (Phillips, 2017). We set the number of
iterations for each FC-RM combination equal to the number of occurrence
(n) for each species thus making it equivalent to n-1 jackknife folds,
which is a good approach for species with small
occurrence data (Warren and
Seifert 2011, Shcheglovitova and Anderson 2013, Radosavljevic and
Anderson 2014, Galante et al. 2018). Crossing 11 RM values and four FC
sets we built 44 sets of models for each species. Maxent generates (n+1)
models including one composite (average) model for each set of RM-FC
combination.