Invasive species are an economic and ecological burden, and efforts to limit their impact are greatly improved with reliable maps based on species distribution models (SDMs). However, the potential distribution of new invaders is difficult to anticipate because they are still spreading with few observations in their invaded habitat. Therefore, an accepted practice in predicting the distribution of invasive species has been to incorporate habitat information from its entire geographic distribution (invaded and native ranges) into SDMs. Yet, this approach, due to niche shifts, niche expansions, and data deficiencies, commonly misrepresents where an invasive species is found in its new range. Here, we use time series records (invasion stages) from 13 invasive plant species in North America to explore the tension between modeling invasive species using global range and invaded range data and to determine if there is a "tipping point" at which one SDM strategy performs better than the other in predicting the ultimate distribution. At the earliest invasion stage, models developed using both invaded range and global occurrences on average performed better and had less variability across species than other model strategies at this stage. However, after as few as 100 observations of an invasive plant had been made, US-invaded range models, on average, outperformed global range models and models that combined occurrences. By building models with global and US-scale predictors, we show that higher performance of invaded range models was in part because of greater data quality at the invaded-range scale. Our work demonstrates that after relatively few observations of an invasive species in its invaded range, it is more accurate to model its potential distribution using only information from the invaded range while disregarding information from other regions. This work develops a robust and comprehensive approach to modeling novel distributions of newly observed invasive species.
