Fig. 1: Our approach for evaluating the multiple predictors of network dissimilarity at large spatial scales.

We used several distance matrices (N × N, where N is the number of local networks in our dataset) as variables in the statistical models. a, b Maps show examples of ecoregions and biomes (colors of shaded areas) represented in our dataset. Points indicate the locations of four network sites used to illustrate how we generated our distance matrices (see Fig. 2 to visualize the locations of all network sites in our dataset). Ecoregion and biome distance matrices were generated using both a binary (shown in the figure) and a quantitative approach (generated by measuring the environmental dissimilarity between ecoregions/biomes; see “Methods”). Because ecoregions are nested within biomes, network sites located within the same ecoregion are always within the same biome, but the opposite is not necessarily true; see, for example, the comparison between network site 1 and network site 3, which involves two ecoregions (Southwest Amazon moist forest and Araucaria moist forest) from the same biome (Tropical & Subtropical Moist Broadleaf Forests). c The human disturbance distance matrix was generated by calculating the absolute difference between local-scale human footprint values around each network site. d–f Spatial distance, elevational difference and sampling-related distance metrics (i.e., sampling methods, hours, months, years, and intensity) were used as covariates in our models to control for distance-decay effects and differences in network sampling. Note that even though we only depict the sampling method distance matrix in f, all sampling-related metrics were used as predictors in the models. g–i We used three different facets of network dissimilarity (i.e., species turnover, interaction dissimilarity and network structural dissimilarity) as response variables (see Network dissimilarity section in “Methods”). j We tested the significance of our predictor variables by employing a combination of Generalized Additive Models (GAM) and Multiple Regression on distance Matrices (MRM). In this analysis, the non-independence of distances from each local network is accounted for by performing 1000 permutations of the response matrix. Ecoregions and biomes were defined based on the map developed by Dinerstein et al.³ (available at https://ecoregions.appspot.com/ under a CC-BY 4.0 license). Bird and plant silhouettes were obtained from http://phylopic.org under a Public Domain license.