Background/Question/Methods Extensive economic and environmental damage has been caused by invasive non-indigenous plant species in many ecosystems worldwide. Particularly, the importation of nursery stock has been a major pathway for the introduction of invasive plants into North America. Because new species continue to be introduced – at an increasing rate – the introduction of additional pests is likely without the imposition of effective screening or risk assessment methods. The extremely high costs associated with some invasive species suggests that the cost of errors made when discriminating invasive from non-invasive species is highly asymmetrical. Specifically, if invasive species mistakenly permitted for introduction are more economically damaging than non-invasive species mistakenly prohibited, risk assessment methods that aim to minimize total classification errors will be economically sub-optimal. A superior protocol for risk analysis would be optimized to minimize costs rather than errors. In this study, we test a method for cost-sensitive risk classification of plant species potentially invasive in the United States and Canada. First, policy relevant criteria were used to distinguish three hierarchically related classes of pest plant species (weeds, state-listed noxious species, and federally-listed noxious species). Then, boosted regression trees were combined with cost-sensitive analyses to predict pest class from readily available ecological and botanical characteristics. Results/Conclusions Independent validation showed that prediction of weeds (75% accuracy) and state-listed noxious species (76% accuracy) from three traits was sufficiently accurate to be cost effective. In contrast, we were not able to develop stable predictors for the relatively small set of federally noxious species. Further, these findings were robust to considerable cost uncertainty. Key predictors of weeds and state-listed noxious species included seed mass and maximum height. Facultative wetland habitat association was also a predictor of weed species, while maximum chromosome number was the third important predictor of noxious species. Graphical decision trees are provided that support application of this model to new species not considered by us. Our results demonstrate that cost-effective cost-sensitive screening for invasive plants can be performed using a small set of traits – for which large, publicly accessible databases are available on the Internet - and their interactions.
