Abstract
<jats:p>Invasive alien species (IAS) represent a major risk to threatened endemic species (TES), particularly on islands, but predicting TES-IAS interactions remains a major challenge. We cross-evaluated the complementarity of three predictive approaches, each being implemented with specific datasets from Réunion Island: (i) Joint Species Distribution Models (JSDMs) capturing inter-specific correlations across 147 forest inventories including 31 TES and 28 IAS, (ii) functional overlap based on leaf, height and seed (LHS) trait data for 20 TES and 25 IAS, and (iii) expert-based data identifying 129 TES-IAS pairs as antagonistic from field observations. Using generalized linear models, we assessed the concordance among these approaches. We found that TES-IAS pairs displayed significantly lower JSDM correlations than the community mean. However, antagonistic pairs identified by experts exhibited positive JSDM correlations, revealing a contradiction between expert perception and model-derived interactions. Functional differences significantly explained variation in both JSDM correlations and expert-based antagonistic status, yet their explanatory power remained small (<6%) and variable across functional distinctiveness indices. Strikingly, while greater functional dissimilarity decreased the probability of being identified as antagonistic by experts (supporting niche overlap and resource competition theories), the same dissimilarity was associated with more negative JSDM correlations, suggesting that JSDMs captured unmeasured environmental filtering rather than direct biotic interactions. These discrepancies likely reflect limitations inherent to each approach: (i) JSDMs seem constrained by omitted environmental predictors and non-equilibrium species distributions, (ii) LHS traits may not fully capture the dimensions most relevant to TES-IAS competition, and (iii) expert knowledge may be subject to perceptual biases. We argue that JSDMs could serve as a powerful integrative platform, combining functional traits and expert knowledge to bridge statistical modeling and community dynamics. However, this transition requires longitudinal plot data, more comprehensive environmental descriptions, and a judicious selection of functional traits better reflecting the mechanisms driving TES-IAS interactions.</jats:p>