Sensitivity of bipartite network analyses to incomplete sampling and taxonomic uncertainty

Bipartite network analysis is a powerful tool to study the processes structuring interactions in antagonistic ecological communities. In applying the method, we assume that the sampled interactions provide an accurate representation of the actual community. However, acquiring a representative sample may be difficult as not all species are equally abundant or easily identifiable. Two potential sampling issues can compromise the conclusions of bipartite network analyses: failure to capture the full range of interactions of species (sampling completeness) and failure to identify species correctly (taxonomic resolution). These sampling issues are likely to co-occur in community ecology studies. We asked how commonly used descriptors (modularity, nestedness, connectance and specialisation (H2')) of bipartite communities are affected by reduced host sampling completeness, parasite taxonomic resolution and their crossed effect. We used a quantitative niche model to generate replicates of simulated weighted bipartite networks that resembled natural host-parasite communities. The combination of both sampling issues had an additive effect on modularity and nestedness. The descriptors were more sensitive to uncertainty in parasite taxonomic resolution than to host sampling completeness. All descriptors in communities capturing less than 70% of correct taxonomic resolution strongly differed from correctly identified communities. When only 10% of parasite taxonomic resolution was retained, modularity and specialisation decreased [~]0.3 and [~]0.1-fold respectively, and nestedness and connectance changed [~]0.7 and [~]3.2-fold respectively. The loss of taxonomic resolution made the confidence intervals of estimates wider. Reduced taxonomic resolution led to smaller size of the communities, which emphasised the larger relative effect of taxonomic resolution on smaller communities. With regards to host sampling completeness, connectance and specialisation were robust, nestedness was reasonably robust ([~]0.2-fold overestimation), and modularity was sensitive ([~]0.5-fold underestimation). Nonetheless, most of the communities with low resolution for both sampling issues were structurally equivalent to correctly sampled communities (i.e., more modular and less nested than random assemblages). Therefore, modularity and nestedness were useful as categorical rather than quantitative descriptors of communities affected by sampling issues. We recommend evaluating both sampling completeness and taxonomic certainty when conducting bipartite network analyses. We also advise to apply the most robust descriptors in circumstances of unavoidable sampling issues. Open Research statementwe provide permanent and open access links to data sources and replication code in Appendix S1.