We built a total of 36 networks of interactions for each method, 12 for each habitat type (Paratudal, Canjiqueiral and Riparian forest). Thus, there were 36 networks for the plant-centred method, 36 matrices for the animal-centred method and 36 for the combined method (plant + animal-centred method). In addition, we built a general matrix for each method including all sampled habitats (one total network for plant-centred method, one total network for animal-centred method and one total network combined both methods), totalling 111 matrices constructed.
We evaluated the sampling completeness of the plant-centred, animal-centred and the combined networks (Chacoff et al., 2012; Vizentin‐Bugoni et al., 2016). For this purpose, each combination of a plant and a pollinator species is taken as “species” and the frequency of each pairwise interaction represents its “abundance” (Vizentin‐Bugoni et al., 2016). We estimated the diversity of interactions using Chao 1 estimator of species richness (Chao, 1984; Colwell & Coddington, 1994). Then, we calculated sampling completeness as the ratio of the observed and estimated richness of interactions (Chacoff et al., 2012). The Chao 1 estimator was computed using the iNEXT package (Hsieh et al., 2014) in R (R Development Core Team, 2019). Sampling completeness of the studied networks are shown in the supplementary material (Table S2; Table S3) and did not differ between the different methods (combined network = 0.68±0.17; plant-centred network 0.64±0.19; animal-centred network 0.61±0.17, χ2 = 20.07, P = 0.17).
We evaluated how species partition their interactions using two quantitative metrics for combined, plant and animal-centred networks. Network-wide specialization was estimated by the H2' index that describes whether species restrict their interactions from those randomly expected based on partners’ availability (Blüthgen et al., 2006). Modularity indices quantify the prevalence of interactions within subsets of species in the community and was calculated using the DIRTLPAwb+ algorithm (Beckett, 2016) using the computeModules function in bipartite package. Network indices can be affected by the number of interacting species and sampling effort (Fründ et al., 2016). Hence, besides using the “raw” modularity and specialization values, we also compared z-score transformed metrics, where the mean value of a metric obtained by multiple randomizations of a network is subtracted from the observed value. Then it is divided by the standard deviation of values across all randomized networks, thus indicating to what extent an empirical observation departs from a random pattern defined by a specific null model (Dormann et al., 2009; Dalsgaard et al., 2017). Here, we used 1000 randomizations for each network, with the Patefield null model, which fixes the network size and the marginal totals, i.e., species richness and species’ total number of interactions, while shuffling interactions randomly (Patefield, 1981). Besides these two network-level indices, we also calculated two species-level indices that capture distinct properties of a species in the network: 1) degree, expressing the number of interaction partners’ that each species is linked to in the network; and 2) species-level specialization d', which quantifies how interaction frequencies of a given species deviate in relation to the availability of interaction partners in the network, with higher values indicating higher specialization (Blüthgen et al., 2006). We calculated all network-related indices with the bipartite package version 2.15 (Dormann et al., 2008) in R.