Accounting for observational uncertainty in interaction networks

Background/Question/Methods

Thoroughly sampling biodiversity is difficult. More difficult still is thoroughly sampling the interactions linking species in complex networks. The additional challenge of inventorying interactions is largely due to the dynamic and often infrequent or ephemeral nature of interactions. While much effort has gone into developing and applying modeling techniques accounting for uncertainty in species occupancy, modeling uncertainty in species interactions has received relatively little attention. By extending statistical techniques from occupancy modeling, we develop a model of species interactions robust to observation error. The model decomposes the probability of recording a specific interaction as a combination of the phenology of each interacting species, the affinity of each species for the interaction, and a term accounting for potential observation error. We apply this model to a dataset on plant-pollinator interactions from a subalpine meadow in the Colorado Rocky Mountains, USA. Plant-pollinator Interaction data were collected weekly during mornings over the entire flowering period, May - September, from 2015 to 2019. We demonstrate the consequences of accounting for observation uncertainty by calculating and comparing various network metrics between model estimated and observed networks.

Results/Conclusions

The empirical data include a total of 4,261 interactions across 836 links between 267 species of animal visitors to 41 species of plants. The full network (aggregated across time periods) showed nestedness and connectance values that are typical of plant–pollinator networks aggregated over similar temporal scales (NODF = 25, connectance = 0.08). Analyzing the empirical data using traditional approaches showed large fluctuations in network metrics throughout the flowering period (e.g., a 70.1% reduction in connectance over 15 days). However, these fluctuations are likely due to observation error as the model estimated a 30% increase in connectance over the same period. The model revealed an asymmetrical pattern in which connectance gradually decreased as throughout the first half of the flowering period, then rapidly increased in the latter half. Traditional estimates of NODF also showed large fluctuations that are likely artifacts of observation error. In fact, the model estimates suggest NODF reached its highest value late in the season and maintained that high value throughout the end of the flowering period. Taken together, our results suggest that accounting for observation error in ecological networks can improve both our predictions for and mechanistic understanding of critical ecological interactions.