Selecting optimal models using expert approach
We visually screened and compared which of the suitable/unsuitable binary maps produced from the optimal models selected using the four sequential approaches were ecologically plausible. If none of the optimal models selected through the four sequential approaches resulted in ecologically plausible binary maps we then moved to the next best choice and continued to screen till we arrived at the best possible ecologically plausible map using expert knowledge (Galante et al. 2018). We qualitatively assessed ecological plausibility by considering the known elevation range and the predicted suitable habitat. We also compared the predicted map with any available literature, such as IUCN distribution maps or authors’ field experience. However, we acknowledge here that given the relatively high number of fish and odonate species modelled, compounded by very poor literature on the species from Bhutan, some of the chosen optimal models could be subjective. We developed the binary maps using the “balance training omission, predicted area and threshold value Cloglog threshold” (hereafter ‘balance threshold’) out of 11 Cloglog thresholds generated by Maxent for each model iteration. Though lower thresholds can overpredict suitable habitat, they are better for species with few occurrence data (Pearson et al. 2007, Radosavljevic and Anderson2014) and can also uncover potentially informative distribution areas (Pearson et al. 2007). We also found using the 10th percentile threshold restricted the predicted suitable habitats around occurrence points, or predicted the whole study area as unsuitable, in some cases (See ‘Predicted Habitat’ in Result) (Pearson et al. 2007, Radosavljevic and Anderson, 2014; Galante et al. 2018). We produced binary maps using ArcGIS10.2.2 (ESRI 2014).