Variation in network metrics
In agreement with our first hypothesis, we found that the generality of the observed networks decreased with increasing temperature (observed slope = -0.4973, P = 0.035; Fig. 2, Table 1) which was consistent with a decrease of the overall network specialization (H2’) along the elevation (Appendix S1, Fig. S4a, Table S1). The variation in these metrics differed significantly from null models, as we found their observed slopes to be outside the 2.5–97.5% quantile interval of the slopes obtained from random networks and SES values were high (generality: 2.5–97.5% quantile interval = (-0.4066, -0.1868), SES = -3.44; overall network specialization in Table S1). We further found a negative relationship between the weighted nestedness and the temperature for empirical webs that was different from null models (observed slope = -0.0102, P = 0.042, 2.5–97.5% quantile interval = (-0.0054, -0.0060), SES = -3.51; Fig. 2, Table S). We found a positive relationship between robustness and temperature (slope = 0.0033, P = 0.023), but the observed decrease in robustness in cold environments was lower than expected from nulls models (2.5–97.5% quantile interval = (0.0057, 0.0082), SES = -5.77), indicating a role of the wiring of interactions in attenuating the decrease in robustness toward colder conditions (Fig. 2). Variation in the number of links per species (Fig. 2, Table 1), the connectance and the trophic niche overlap of orthoptera (Appendix S1, Fig. S4b and S4c, Table S1) were not different between the observed and the randomized networks. We found a positive relationship between temperature and the number of links per species in empirical networks (Fig. 2a, Table 1) but not for the connectance or the niche overlap (Appendix S1, Fig. S4b and S4c, Table S1).