Quantitative specialization measures in interaction networks and whether they can predict functional consequences across diversity levels

Background/Question/Methods

Multiple concepts of specialization exist across ecological sub-disciplines and applications. Specialization drives non-neutral structure of ecological networks and is crucial for biodiversity-functioning relationships. There are numerous suggestions of how to quantify specialization, often implying different functional consequences. We compare specialization concepts and quantify the extent to which different specialization metrics reflect different functional consequences. We first simulated plant-pollinator interactions with a quantitative niche model that produced a wide range of specialization among networks. We then calculated functional consequences of specialization with a set of models. The consequences we consider are (1) effective number of hosts or partner species, (2) robustness to secondary extinctions, (3) proportion of available resources used, (4) niche partitioning as reduced interspecific competition, (5) functional complementarity as loss in functional coverage with loss of diversity and (6) pollination benefit of specialist pollinators mediated by conspecific pollen delivery. We also simulate subsampling of interaction networks, calculate eight popular specialization metrics and compare them across different sampling intensities and to the theoretical functional consequences. We thus bridge a gap between empirical data with limited information, null models that do not include specialization and theoretical models of functional consequences poorly connected to real data.

Results/Conclusions

We show that

(1) Network size (consumer and resource diversity) strongly influences the relationship between the most basic definition of specialization (generality) and functional consequences. This highlights that it is fundamentally impossible to reflect all specialization consequences across networks of variable size with a single metric.

(2) With realistic sample sizes, all metrics are biased towards specialization. Importantly, this bias depends on the underlying degree of specialization, being strongest for generalized networks. The severity of sampling bias strongly differs among metrics. We provide a guideline for which metrics to use until better methods are available.

(3) When all species have the same number of observations, all metrics tend to reflect underlying specialization despite sampling bias. However, different metrics are good for different purposes and not all are useful for variable network size. Metrics often reflect about 50% or less of variation in functional consequences. The limitation might be more severe under real-world scenarios with less extreme variation in specialization than in our simulations.

Studies that need to quantify specialization should carefully consider their objectives, relevant concepts and sampling biases. This will help determine best metrics and ultimately interpreting network data with respect to functional consequences for species, communities and ecosystems.