Simulate networks to understand sampling artefacts that obstruct the measurement of ecological specialization

With a little delay caused by my moving back to Germany, I’d like to to advertise my most recent paper (available in Oikos, includes a rich online appendix), which I find really important:

The network approach is currently popular in ecology, as it promises to manage the complexity of interactions among many different species. When network studies try to include all interacting species, some of them will inevitably be represented by few observations. Specialization, a major aspect of network structure, is overestimated with few observations: there is a sampling bias (in the sense of “bias due to sampling”). Trying to understand the functional relevance of biodiversity and the impact of environmental change on communities and ecosystems, more studies now compare different networks, which further limits how much effort can be spent to sample each network.

In this study, we developed a model that generates realistic quantitative interaction networks and used it to evaluate methods that try to overcome sampling bias in specialization estimates. We found that, unfortunately, all metrics and methods currently used for network analysis misrepresent true network structure when used on data with realistic numbers of observations. Although some metrics performed reasonably well, caution should be used when comparing empirical network patterns to theoretical predictions and when comparing different networks. Our model could be useful for carefully evaluating the potential for sampling bias for a given study and develop new methods to correct quantitative estimates of network structure. Our study also highlights the large potential for sampling bias in studies estimating specialization without a network focus.