Using empirical stressor-response models to make predictions at finer spatial scales

Empirical stressor-response models at the regional or national scales have utility for making inferences at the same spatial scale used to develop the models.   We would like to use these models to predict whether effects would be expected at specific sites or within smaller areas.  To be useful for this purpose, the relationships observed at coarser scales must hold for smaller subsets of the data.   We applied national models that predict amphipod toxicity test results from sediment chemistry to regional subsets of the data.   Although in some circumstances the national model fit very well, in other cases, there was little relationship between the national model predictions and the regional data.   We investigated several explanations for these results.  When explanatory variables are correlated, relationships may not hold because of differing covariation patterns within smaller subsets of the data.  However, chemical covariation patterns did not predict how well models performed.  We also investigated the influence sediment particle size (i.e., % fines), and the benefits of identifying and eliminating individual chemical predictors that produced high false positive rates.  Our results underscore the need to evaluate the performance of empirical stressor-response models with subsets of the data and to calibrate them with site- or region-specific data prior to application to new data sets.   Disclaimer:  The views expressed in this presentation are those of the authors and do not necessarily reflect the views or policies of the U.S. Environmental Protection Agency.