EFFECTS OF SAMPLING BIAS ON ROBUSTNESS OF ECOLOGICAL METRICS IN FOSSIL PLANT-DAMAGE TYPE ASSOCIATION NETWORKS

Researchers have long used insect damage on fossil leaves to understand the structure of associations in fossil plant–insect herbivore assemblages. The codification and standardization of damage types (DTs), has helped us to comprehend major patterns within and among fossil and modern assemblages. The use of bipartite networks to quantify these complex systems facilitates measurement of fine-resolution properties which are not possible using traditional aggregated metrics, such as DT diversity. Network-level metrics, which measure aspects of ecological specialization, co-occurrence, and nestedness, can then be used for more in-depth cross-assemblage comparisons than is possible with currently available metrics.

Sample sizes in fossil data can vary considerably among systems for a variety of reasons, including differences in taphonomy and collection efforts. As network metrics depend upon the exact structure of interactions, differences in sample size among assemblages can affect the validity of comparisons using these metrics. As a result, comparisons across systems might reflect patterns in data collection process instead of actual ecological differences, a problem also true of traditional DT-based metrics. To test such effects and assess the robustness of various network measures in plant-DT networks, we analyzed 60 angiosperm-dominated floras with varying sample sizes and plant diversities. We examined changes in the values of network metrics through subsampling procedures. We found that network metrics are differentially sensitive to issues of sampling, but that some of the metrics are reasonably robust to these simulated processes of data loss and reconfiguration. Better performing network metrics, such as NODF, H2, connectance, and niche overlap among others, were consistent across sampling intensities, allowing quantification of their robustness and consistency. Our efforts also provide a general quantitative framework for accurate comparisons among assemblages of varying sample sizes using network metrics.

We also utilized models of Bayesian to infer plant-DT association network structure from observed noisy fossil data. This approach calculates uncertainty about the structure of our constructed network, while taking collection bias into account. Such estimates of ‘association certainty’ can identify which components of a fossil interaction network are known with confidence and also help identify specific associations that would benefit from a better sampling effort or expert knowledge.