Moderators (predictors)
We investigated different categories of moderators, including absolute latitude and elevation, as well as climate, productivity (annualPET), topography (TRI, topoWET), leaf size and average temperature during the experiments. Additional local moderators included available potential prey (abundance of herbivores) for predators and abundance of predators on experimental leaves for herbivores.
To create the climate gradients, we projected separately precipitation and temperature variables in multivariate space using principal component analysis (PCA) (Romero et al ., 2018). Scores of the first axis of PCA for temperature variables (PC1temperature, which explained 78% of the total variance) represented a gradient spanning from higher temperature variability to warming and lower temperature variability. Conversely the first axis of PCA for precipitation variables (PC1precipitation: 64% of the total variance) represented a rainfall (dry to wet) gradient, at global (Fig. 1) and local scales (Figs. S2-S3). PC1precipitation included mean annual precipitation (bio12), precipitation of driest month (bio14), index of the degree of water deficit below water need (aridity) and precipitation seasonality (coefficient of variation) along the year (bio15) (Table S1). The only variable denoting rainfall variability (bio15) did not contribute much to the PCA1 variance (Fig. 1, Fig. S2). Therefore, the PC1precipitation represented a rainfall gradient, which varied from dry (negative values) to wet conditions (positive values) (Fig. 1, Fig. S2). Sites defined as being climatically more suitable were those characterized by higher PC1precipitation scores. PC1temperatureincluded mean annual temperature (bio1), mean diurnal range (bio2), temperature seasonality (bio4) and temperature annual range (bio7). The signs of original PC1temperature scores were multiplied by -1 to change their graphic presentation for the ease of interpretation, matching with precipitation trends, thus producing a gradient from temperature variability (negative values) to warming and lower temperature variability (positive values). Sites defined as being thermally more variable were those characterized by higher intra-annual (e.g., bio4, bio7) or diurnal (e.g., bio2) variation in temperature (Fig. 1, Fig. S3).
Some moderators were collinear (Fig. S4), and were thus removed from the models based on their variance inflation factor (VIF) (Zuur et al., 2010). The removed moderators were Latitude, topoWET and AnnualPET. All remaining moderators had VIF < 2. The best-fitting linear models were then obtained through backward selection; the best models retained were those to present the smallest Akaike Information Criteria corrected for small sample sizes, AICc (∆AICc>2). Null models (intercept-only effects, i.e., no moderators included) were contrasted with models containing moderators plus the intercept.