Effects of modelling choices on early invasive risk assessments: lessons from the psyllid vectors of Citrus Huanglongbing

Abstract Species distribution models (SDMs) are widely used in early invasion risk assessments and are increasingly critical for anticipating biological threats under global change. Yet, SDM outcomes can vary substantially depending on modelling choices, many of which remain poorly standardized. These variations have tangible consequences when model outputs guide regulatory and management decisions. Here, we systematically evaluated how five key modelling decisions affect the performance and spatial projections of early-stage invasion risk models, including strategies that reflect the most common practices: (i) quality and curation of occurrence data, (ii) use of native-only versus native and invasion occurrence data in model calibration, (iii) background point selection, (iv) environmental variable selection, and (v) ensemble versus single-model approaches. In addition to conventional whole-range evaluation metrics (AUC, TSS, Boyce index), we applied a complementary index focused on the invaded range: the invasive False Negative Rate (iFNR). Finally, differences in projected areas at risk were estimated using the Jaccard index of overlap. As a case study, we modelled two invasive insects with contrasting invasion histories and distributions — Diaphorina citri and Trioza erytreae — vectors of the bacteria responsible for Huanglongbing (HLB), the most devastating citrus disease worldwide. Our results demonstrate that expert-curated occurrence data and inclusion of validated invaded-range records consistently improved model performance in novel environments. Moreover, conventional classification metrics alone failed to distinguish between models producing widely divergent invasion forecasts. Incorporating invasion-focused metrics and comparing spatial projections was essential to discriminate between models and to support more robust, policy-relevant decisions in invasion risk assessments.