Catching the effects of biotic interactions on community data: partial correlations outperform marginal ones with proper abiotic modelling.

A major goal of community ecology lies in the deciphering of the processes underlying species distribution. A widespread approach to this question is to identify patterns in species community data and relate them to possible processes. Joint Species Distribution Models (JS-DMs) offer one way to do so through the infernece of association networks that describe patterns of statistical correlations and dependencies between species, but it is unclear what processes can explain the presence of such correlations. While it has now been established that there is no equivalence between JSDM-inferred associations and biotic interactions, the later remain one possible explanation, among others, for the former. However, to our knowledge, there is no specific study of the statistical patterns induced by different types of interactions or of the conditions under which they may or may not appear as statistical correlations / dependencies in species communities. To explore these questions, we propose a "virtual ecologist" approach that consists in simulating community data based on abiotic and biotic processes with the VirtualCom model that emulates the effects of environmental processes and of competition and facilitation interactions. Then, we study to what extent JSDMs retrieve correlations between species that match the simulated interactions. We show that these interactions are better identified when using JSDMs that model partial correlations between species rather than marginal ones. We further demonstrate how critical it is to correctly model abiotic effects in order to identify biotic ones and that the "correct modelling" of these effects depend on the type of interactions at stake.