The accuracy of species’ range mapping is important for providing reliable baselines to monitor change. Range maps are often based on habitat models built from large datasets containing  presence/absence data. The largest forest monitoring data set in the United States is from the Forest Inventory and Analysis Program (FIA) of the U.S. Forest Service. FIA recently switched from a mix of plot designs optimized for specific areas and owners to a nationally standardized design; we need to determine the extent that different tree selection probabilities emerging from different designs affect model development. Here, we select and compare models built from data segregated by sample design for twenty-five tree species residing in the western United States. We build species-habitat models based on climate using Non-Parametric Multiplicative Regression (NPMR), an approach designed for complex responses typical of species-habitat relations. Model selection for non-parametric, machine-learning approaches such as NPMR is a field of active research. We perform model selection by using generalizability to unseen data as the dominant criterion. 

Results/Conclusions We find that a majority of species yield similar models with shared predictors. However, the models show some differences in fits and climate domains expressing statistical interactions. We discuss the implications for range mapping applications.