Statistical analysis
We contrasted plant–pollinator network indices according to sampling method using linear mixed-effects models in the lme4 package in R (Bates et al., 2015). We took network indices (modularity and specialization) and sampling completeness as response variables, network type (plant-, animal-centred or combined network) as fixed- and plots identity as random-factors in the models. The statistical significance was tested using a likelihood ratio test (Anova type II) comparing the model with and without the fixed factor using the R package car (Fox & Weisberg, 2011). For network dissimilarities, models were constructed considering β-plants, β-pollinators, network beta diversity of interactions (βWN), species turnover (βST), rewiring (βOS) and "true" beta diversity which accounts for richness differences (βWN.repl, βOS.repl) as response variables, network type (plant-, animal-centred or combined network) as fixed- and plots identity as random-factors.
We contrasted the species level indices (degree and specialization) between the methods for both plants and pollinators, taking these indices as response variables, network type as fixed factor and identity of species and plots as random factors in the models. After detecting significant factors in the models, we conducted multiple comparisons (post hoc Tukey test) using the function glht in the package multcomp (Hothorn et al., 2008). In addition, we compared the contribution of turnover (βST) and rewiring (βOS) to the beta diversity of interactions (βWN) and compared β-plants and β-pollinators between sites using a t test.
In order to understand how plant species level indices (degree and specialization) calculated for the combined network are determined by plant traits, we performed linear models including plant species traits: 1) floral type (flag, gullet, dish, inconspicuous (small pale flowers), brush and tube); 2) flower size (width and length); and 3) pollination system: small bees, large bees, oil bees, butterflies (including moths), beetles and “dsi” (diverse small insects) (Frankie et al., 1983; Machado & Lopes, 2004; Danieli-Silva et al., 2012) (Table S1). We used floral types as a trait to measure this difference between sampling methods, once flower openness indicates the accessibility to flower resources by pollinators (Olesen et al., 2007). Since we only sampled one species with beetle pollination system (Annona nutans), we removed it from the analysis. We included all plant traits as fixed effects, and species level indices as response variables.
We also built linear models using the absolute difference of degree and specialization between the plant-centred and combined network as response variables (∆specialization d' and ∆degree), to evaluate how plant traits modulated the effect of adding pollen data on estimates of species specialization. Then, for significant fixed effects we conducted multiple comparisons with posthoc Tukey test using the function glht in the package multcomp (Hothorn et al., 2008). All statistical analyses were conducted in R (R Development Core Team, 2019). All models built were checked for normality using the uniformity test (Kolmogorov-Smirnov test), dispersion of the residuals and spatial autocorrelation (Moran's I test) using the DHARMa package (Hartig, 2020). After detecting deviation from model assumptions, we log10 transformed species degree, and the turnover component of beta diversity (βST) as well as plant and pollinators β diversity in the models comparing dissimilarities within and between habitats.