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

In simulated communities with known competition and facilitation networks, JSDMs recover only modest interaction signals, with best performance for shortlist ranking under specific model integration and data conditions.

• Why it matters: It is unclear how much of true biotic interaction structure an estimated JSDM association network retains versus what comes from shared environmental niches alone. This matters for interpreting inferred species networks as candidate interactions.
• What they did: A virtual ecologist simulated communities along an environmental gradient with known interaction networks under three mechanisms: VirtualCom lottery for space, Gompertz, and Ricker. Four estimators with increasing integration were compared, including a Gaussian graphical model on log-transformed counts and PLNnetwork, and each inferred network was scored for ranking and shortlist accuracy against paired simulations without interactions.
• The result: Partial associations beat marginal ones, and JSDM gains depend on use and mechanism: count-model integration improved ranking under Gompertz, while jointly estimating a sparse network improved shortlists under all three mechanisms. Performance required near-error-free covariates, no unmeasured gradient, a suitable abiotic covariate representation, and enough latent factors, and inferred networks were at best candidate interaction shortlists.

AI summary of the abstract · may contain errors · not medical advice · read the paper.