PS 81-185 - Model application niche analysis: An approach for assessing the transferability and generalizability of ecological models

Friday, August 11, 2017
Exhibit Hall, Oregon Convention Center
Jessica B. Moon1, Theodore H. DeWitt2, Melissa Errend3, Randy Bruins4, Mary E. Kentula5, Sarah Chamberlain6, M. Siobhan Fennessy7 and Kusum J. Naithani1, (1)Biological Sciences, University of Arkansas, Fayetteville, AR, (2)NHEERL Western Ecology Division, U.S. EPA, Newport, OR, (3)College of Earth, Ocean, and Atmospheric Sciences, Oregon State University, (4)US EPA, (5)NHEERL Western Ecology Division, U.S. Environmental Protection Agency, Corvallis, OR, (6)Department of Geography, The Pennsylvania State University, (7)Biology Department, Kenyon College, Gambier, OH

A 30-year review on predictive models used in regulatory decision-making, revealed that transferring models to contexts other than that for which the models were developed was one of the biggest vulnerabilities to their legal defensibility. The use and transfer of models by ecologists, to inform environmental management and decision-making, has grown exponentially in the past 50 years. Given this trend, and the importance of public confidence in the decisions that are being made based on models, model users need better ways to evaluate the possibility of misuse when transferring models to new contexts. We present one approach, a model application niche analysis, where ecologists synthesize information from databases and past studies to create model performance curves and decision landscape plots. These visualization tools characterize a model’s application niche as a function of model performance and uncertainty across dimensions of context, as a means to evaluate both model transferability and generalizability.


To demonstrate the utility of this approach, we evaluated an empirical model developed to predict the mean coefficient of conservatism (i.e., ecological condition) of plant communities in wetlands across Pennsylvania for transfers across the contiguous U.S. The model predictors include surface soil pH and the landscape development index, a measure of anthropogenic activity in the wetland’s surrounding landscape. Using model performance curves and decision landscape plots along latitude and elevation gradients, we show that this model is transferable to locations in the eastern, southeastern and Pacific Northwest regions of the U.S. where high forest cover, rainfall and geology contribute to acidic soil formation. We also show that model performance is weakest, and the model structure is questionable, in regions of the U.S. with basic soil conditions, such as the west and the tips of Florida and Louisiana. Our goal is to invoke further inquiries into the development of consistent and transparent practices for model selection when transferring ecological models.