Species distribution models are rapidly proliferating in studies relating to climate change and the spread of introduced species. These modeling approaches have clear merit, but pitfalls exist that can bias predictions. These limitations include the use of presence only data, insufficient sampling of environmental parameter space, inclusion of variables at the incorrect spatial resolution, models that are not tested with independent data, and those that predict distributions beyond known parameter space. To quantify how these problems can bias model predictions, we use presence and absence data for Argentine ants (Linepithema humile) independently collected in two regions of southern California. First, we demonstrate that insufficient sampling of environmental parameter space incorrectly predicts species distributions in novel regions. Second, models using data that sufficiently sampled the environmental parameter space resulted in the best and most general models. Lastly, environmental variables differed in importance across spatial grain; the best multi-variable models at 100-m resolution differed from those at 10-km resolution. Taken together, our results suggest that caution should be used when projecting model predictions into novel environments and the importance of using variables that are meaningful with respect to the spatial resolution of the data being analyzed.