The effect of biotic factors on overall model performance
The models improved significantly when the interacting species was included, for all evaluation measurements: Area Under the Curve (AUC), continuous Boyce index (CBI) and corrected Aikake information criteria (AICc). The difference in these evaluation metrics is the change from models with biotic interactions to models without biotic interactions. The interacting species was included at a resolution of 1 km, which was on average the optimal resolution for most models (fig. 2A). The models of the cleptoparasitic, oligolectic and polylectic bees all showed a statistically significant increase in evaluation AUC (fig. 3A), evaluation CBI (fig. 3B) and a decrease in calibration and evaluation AICc (fig. 3C; Appendix S5). There were not enough samples in the evaluation data to calculate the AICc for a proportion of the bees (34.02% of the species; 66 species) and only the cleptoparasitic bees showed a significant improvement on the species and genus level (fig. 3D). The biotic interaction had a high contribution to the models of all the functional bee groups relative to the climate, land use and soil variables (fig. 4; Appendix S6). The biotic interaction had the highest contribution to the models compared to the climate, land use and soil variables for the cleptoparasitic bees.
The models of the cleptoparasitic bees showed a statistically significant higher improvement in evaluation AUC, evaluation CBI and in calibration and evaluation AICc compared to the oligolectic and polylectic bees with the interacting species at both taxonomic levels (Appendix S5). The difference in evaluation metrics between oligolectic and polylectic bees was in no case significant.
We found that the relative performance of the interacting species compared to any randomly chosen interaction was the highest for the cleptoparasitic, oligolectic and polylectic bees respectively and the polylectic bees showed a more general preference of biotic interactions than the cleptoparasitic and oligolectic bees. The models with the interacting species (added at species and genus taxonomic level) had an evaluation AUC that scored within the highest 5% best performing models of the nullmodels, i.e. the models with the random interactions, in 52.8% and 62.3% of cleptoparasitic bees, 36.4% and 43.2% of oligolectic bees and 13.4% and 14.4% of polylectic bees (fig. 5; Appendix S3). When the threshold was within the 25% best performing models, these percentages increased up to 83% and 94% of the cleptoparasitic bees, 56.8% and 63.6% of the oligolectic bees and 40.2% and 51.2% of the polylectic bees. The polylectic bees showed a less skewed distribution, but more of a continuous distribution, suggesting a less selective response to the interacting species and a more general preference of biotic interactions (fig. 5A).