Ecological network inference is not consistent across scales or approaches

Abstract

Several methods of ecological network inference have been proposed, but their consistency and applicability for use across ecologically relevant scales require further investigation. Here, we infer ecological networks using two data sets (YFDP, FIA) describing distributional and attribute information at local, regional, and continental scales for woody species across North America. We accomplish this inference using four different methodologies (COOCCUR, NETASSOC, HMSC, NDD-RIM), incorporating biological data along an occurrence-performance spectrum while accounting (or not) for various confounding parameters. We contrast 1-1 associations at each evaluated scale to quantify consistency amongst inference approaches. We also assess consistency across scales within each inference approach. Ultimately, we find that inferred networks are inconsistent across scales and methodologies, particularly at continental scales. We highlight how such inconsistencies between network inference methods may be linked to using occurrence or performance information and incorporating or not confounding factors. Finally, we argue that identifying the "best" inference method is non-trivial. Thus, to facilitate the choice of inference methods for a given purpose, we suggest aligning specific research questions and the scale applicability of the method when interpreting the inferred links, network topology, and ecological processes governing network assembly.

A Novel Simulation Framework for Validation of Ecological Network Inference

“…Aligning with previous recommendations for choice of network inference approach , we find that network inference accuracy is governed by (1) spatial biodiversity input type along a spectrum of co-occurrence, abundance, and performance which confers increasing accuracy onto networks inferred therefrom, and (2) environmental considerations in network inference approaches with network inference approaches which account for environmental conditions rendering more accurate inferences of true networks than those approaches who don't. These findings confirm previous criticisms of co-occurrence based network inference approaches failing to capitalise on existing information in abundance and performance data (Bimler et al, 2022;Kusch et al, 2023;Ovaskainen et al, 2017). In addition, our findings also highlight the previously suggested fallacy of neglecting environmental gradients in network inference leading to erroneous assignment of positive and negative because of differences in environmental preferences rather than actual interactions (Blanchet et al, 2020).…”

“…We carried out ecological network inference using two approaches – HMSC and COOCCUR. We chose these two approaches for their unique placements along a co-occurrence to performance spectrum which characterises network inference input data types and which has been demonstrated to affect network inference outcomes (Kusch et al, 2023). While COOCCUR incorporates co-occurrence data, HMSC can consider species performance implicitly through abundance products as well as explicitly through, for example, population growth rates.…”

“…Thus, practitioners of ecological network inference need to carefully consider which network inference approach to use as inferred networks differ considerably across scales and approaches. To streamline choice of inference methodology for specific study purposes, criteria of inference approach applicability have been proposed (Kusch et al, 2023). However, objective measurements of inference performance to judge quality of ecological network inference remain absent.…”

“…A recent comparison of network inference outcomes has demonstrated that – even when analysing the same ecosystem – inferred networks are inconsistent across inference approaches and geographic scales (Kusch et al, 2023). Thus, practitioners of ecological network inference need to carefully consider which network inference approach to use as inferred networks differ considerably across scales and approaches.…”

“…Thus, a simulation framework for network inference validation must explicitly consider bioclimatic preferences of ecological network-contained species to represent more realistic biodiversity products for network inference. Third, spatial biodiversity products for network inference beyond co-occurrence records (i.e., abundance or performance matrices) have been demonstrated to lead to vastly different inference outcomes even when using the same underlying statistical framework for network inference (Kusch et al, 2023). Conclusively, to quantify the accuracy-changes in network inference owing to differences in spatial input types, the simulation framework must be capable of generating biodiversity products beyond co-occurrence.…”