When can a joint species distribution model capture biotic interactions? A simulation study across interaction mechanisms and levels of model integration

Joint species distribution models (JSDMs) describe the composition of communities by the response of each species to the environment and by the residual dependence between species, which can be read as a network of associations. These associations are not biotic interactions, but interactions are one of the processes that can produce them, and it remains unclear how much of a set of interactions an estimated network retains, and under which conditions. We address this question with a virtual ecologist approach. Communities are simulated along an environmental gradient with a known network of competition and facilitation, under three interaction mechanisms: a lottery for space (the VirtualCom model) and two models of population dynamics, Gompertz and Ricker. We compare four estimators of increasing integration, from a Gaussian graphical model on log-transformed counts to PLNnetwork, a Poisson log-normal model that estimates a sparse network of partial associations while it is fitted. Each estimated network is scored for two uses: ranking all the pairs of species, and shortlisting a fewpairs. Each score is compared with the one obtained on a paired simulation without interactions, so that what an estimator owes to the niches alone is not counted. Within the assumptions of these simulators, five results emerge. (i)When the niches are narrow, interacting species seldom meet and there is little to capture, whatever the estimator. (ii) Partial associations do better than marginal ones, mostly among the best-ranked pairs. (iii) What a JSDM brings depends on the use and on the mechanism: its model for counts improves the ranking of all the pairs under the loglinear mechanism (Gompertz) only, mostly when counts are small or sampling effort is unequal, whereas estimating the sparse network together with the model improves the shortlist under the three mechanisms. (iv) Under the lottery and the Ricker mechanism, the Gaussian graphical model does as well over the whole ranking. (v) These gains require a covariate measured almost without error, no unmeasured gradient, a suitable representation of the abiotic effect of the covariate in the model, and enough latent factors. The scores remain modest throughout: a network inferred by a JSDM is at best a shortlist of candidate interactions, under conditions that this study makes explicit.