Sampling biases across interaction types affect the robustness of ecological multilayer networks

Ecological communities rely on complex networks of species interactions. While traditional studies often focus on single interaction types (e.g. plant-pollinator or host-pathogen), there is growing recognition of the need to consider multiple interaction types to accurately model community dynamics. Multilayer networks can be used to model multiple interaction types simultaneously, but building them poses challenges due to the different sampling techniques and expertise needed for documenting different interaction types. This can introduce biases that affect the completeness of data across layers (interaction types). The extent to which such biases affect multilayer network properties remain unclear. Here, we explored this issue using empirical interaction data collected through standardized field sampling in three archipelagos along a latitudinal gradient (the Balearic, Canary, and Galapagos islands). Based on these observations, we compiled three multilayer networks, each incorporating three types of plant-animal interactions: plant-pollinator, plant-herbivore, and plant-seed disperser. We then enhanced these networks by adding interactions from the literature. The observed and enhanced multilayer networks were compared to evaluate how the quantity and bias of missing information affected network properties. In the enhanced networks, the number of herbivore, pollinator and seed disperser interactions exceeded those from the observed networks by, on average, 82 %, 62 % and 96 %, respectively. The species present in the enhanced networks but missing in the observed networks exhibited distinct structural properties. These sampling biases affected both static and dynamic network properties, and the effects varied notably across archipelagos. Observed networks from the Balearic and Canary Islands were less robust to plant removal than their enhanced counterparts, while the opposite was true for the Galapagos. This study, the first to examine the effects of sampling bias on inferred robustness of ecological multilayer networks, reveals that missing data can have complex, hidden effects on modelled network dynamics. Missing data could, therefore, have important implications for predicting and mitigating species loss.