Effect of inventory method on niche models: Random versus systematic error

CrossMark ARTICLE INFO ABSTRACT Article history: Data from large-scale biological inventories are essential for understanding and managing Earth's ecosystems. The Forest Inventory and Analysis Program (FIA) of the U.S. Forest Service is the largest biological inventory in North America; however, the FIA inventory recently changed from an amalgam of different approaches to a nationally-standardized approach in 2000. Full use of both data sets is clearly warranted to target many pressing research questions including those related to climate change and forest resources. However, full use requires lumping FIA data from different regionally-based designs (pre-2000) and/or lumping the data across the temporal changeover. Combining data from different inventory types must be approached with caution as inventory types represent different probabilities of detecting trees per sample unit, which can ultimately confound temporal and spatial patterns found in the data. Consequently, the main goal of this study is to evaluate the effect of inventory on a common analysis in ecology, modeling of climatic niches (or species-climate relations). We use non-parametric multiplicative regression (NPMR) to build and compare niche models for 41 tree species from the old and new FIA design in the Pacific coastal United States. We discover two likely effects of inventory on niche models and their predictions. First, there is an increase from 4 to 6% in random error for modeled predictions from the different inventories when compared to modeled predictions from two samples of the same inventory. Second, systematic error (or directional disagreement among modeled predictions) is detectable for 4 out of 41 species among the different inventories: Calocedrus decurrens, Pseudotsuga rnenziesii, and Pinus ponderosa, and Abies concolor. Hence, at least 90% of niche models and predictions of probability of occurrence demonstrate no obvious effect from the change in inventory design. Further, niche models built from sub-samples of the same data set can yield systematic error that rivals systematic error in predictions for models built from two separate data sets. This work corroborates the pervasive and pressing need to quantify different types of error in niche modeling to address issues associated with data quality and large-scale data integration.