The elusiveness of biotic interactions in spatial data

Joint species distribution models may not yet be able to detect the signal of biotic interactions from empirical community data … due to the lack of sufficiently dense ecological datasets and fast-and-accurate algorithms.

Above: Eurasian nuthatch (Sitta europaea) at its nesting site in a former woodpecker cavity.
(Photo by Josefine S. / CC BY-NC-ND 2.0 /

Every passionate naturalist knows how ecological communities are shaped by biotic interactions. Predators control the abundance of prey populations. Species at the same trophic level compete fiercely for resources. Other seemingly unrelated species form tight relationships of mutual benefit. Yet, when we describe ecological systems at scales above the very local, we usually neglect the effects of these interactions and assume that the environment is the prime determinant of ecological variation. But how much of an oversimplification is this? Theoretical consideration and simulation studies indeed suggest that the signal of biotic interactions should vanish at coarser spatial resolutions, but few studies have tested this proposition empirically. Thus, the aim of our paper “Scale dependency of joint species distribution models challenges interpretation of biotic interactions” was to fill this gap.

Editors’ Choice article: (Free to read online for a year.)
König, C., Wüest, R.O., Graham, C.H., Karger, D.N., Sattler, T., Zimmermann, N.E. and Zurell, D. (2021), Scale dependency of joint species distribution models challenges interpretation of biotic interactions. J Biogeogr. 48:1541–1551.

The main idea originated at least six years ago, when the question of how to account for biotic interactions in species range predictions took up more and more pace. Joint species distribution models (JSDMs) had just come up as a new tool in spatial ecology and the prospect of disentangling the environmental and biotic drivers of species’ ranges was quite exciting. In contrast to classical single-species distribution models, JSDMs simultaneously model the environmental response of multiple species in a community. This joint approach allows us to look not only at species-environment relationships, but also at the residual structure in their (co-)occurrences that is not accounted for by the environment. The general idea underlying JSDMs is to statistically describe this residual structure and derive coefficients for pairwise species associations from it. These species associations (sometimes also called residual correlations) should then tell us whether a given pair of species co-occurs more or less often than expected by their environmental responses, and thus might be indicative of a positive or negative biotic relationship between those species.

High elevation forest habitat in the Swiss Alps.

However, already in very early discussions with collaborators from the fields of macroecology, statistics and ornithology, we were wondering about potential mismatches between the local scale at which interactions take place and the (often coarser) scale at which species occurrence data are available. If JSDMs were indeed able to separate biotic from abiotic signals in occurrence data, we hypothesized, scale mismatches should lead to a systematic change in JSDM estimates across different spatial resolutions. For example, two species might compete for nesting sites at the local scale while still preferring the same habitat overall, which should lead to a negative association at fine resolutions and a positive one at coarse resolutions. Initial simulation studies supported this intuition.

We used a very rich, long-term dataset of Swiss breeding birds from our collaborators at the Swiss Ornithological Institute Sempach. In this dataset, every breeding territory across a grid of more than 250 survey sites is marked during three visits per year, allowing us to define bird communities at varying spatial resolutions within a survey site. We benchmarked a few different JSDM implementations on the dataset and eventually settled for the one with the best balance between statistical flexibility and runtime efficiency. Nonetheless, the combination of multivariate, multi-level Bayesian models and large, long-term data turned out to be  computationally quite challenging. The models at the finest spatial resolutions had a runtime of almost two weeks on a high-performance cluster, and every change in the methodology or model specification would require another two weeks of data wrangling.

However, once we had set up the models correctly and the MCMC algorithm did its magic, we were excited to analyse the results. To our surprise, they were not exactly as expected. Although we did find a moderate shift towards higher estimates of pairwise associations at coarser spatial resolutions, the majority of values were well above zero, indicating a positive spatial relationship among most species at most observational scales. Moreover, the estimates for a given species pair changed rather erratically from one resolution to another, so how exactly would you pinpoint the scale, at which estimates of species association accurately reflect a biotic interaction? We tried to do that by comparing the JSDM estimates to an independently derived matrix of pairwise functional similarity, assuming that functionally similar species should compete more strongly for resources and, thus, tend to have more negative values of species associations. Once again, our results surprised us by showing the exact opposite pattern: species with similar traits tended to have more positive species associations, especially at finer grain sizes. Overall, these results strongly suggested that JSDMs were not able to detect pairwise interactions among Swiss breeding birds in the analysed dataset, but rather that estimated species associations reflected common responses to unmeasured environmental gradients.

Although our findings challenge the notion that JSDMs can detect the signal of biotic interactions from empirical community data, we still think the underlying statistical reasoning is solid and, in principle, would be up for the task. The challenges seem to lie more in the collection of sufficiently dense ecological datasets and the implementation of sufficiently fast and accurate algorithms to deal with them. The ecological community is working hard to make progress on these fronts and we are thus hopeful that the JSDM approach will eventually help solving the elusiveness of biotic interactions in spatial data.

Written by:
Christian König & Damaris Zurell
Institute for Biochemistry and Biology, University of Potsdam, Potsdam, Germany

Published by jbiogeography

Contributing to the growth and societal relevance of the discipline of biogeography through dissemination of biogeographical research.

Leave a Reply

%d bloggers like this: