You are here:
Deriving Criteria-supporting Benchmark Values from Empirical Response Relationships: Comparison of Statistical Techniques and Effect of Log-transforming the Nutrient Variable
Citation:
TREBITZ, A. S. Deriving Criteria-supporting Benchmark Values from Empirical Response Relationships: Comparison of Statistical Techniques and Effect of Log-transforming the Nutrient Variable. Freshwater Science. North American Benthological Society, Lawrence, KS, 31(3):986-1002, (2012).
Impact/Purpose:
To document research results.
Description:
In analyses supporting the development of numeric nutrient criteria, multiple statistical techniques can be used to extract critical values from stressor response relationships. However there is little guidance for choosing among techniques, and the extent to which log-transformation of endpoints affects technique applicability and the critical values obtained is poorly appreciated. This study elucidates these interacting aspects of the analysis process, through side-by-side evaluations of 6 techniques (logistic regression, conditional probability analysis, linear regression, quantile regression, piecewise regression, CART) using data describing algae, water clarity, vegetation structure, and fish composition across a nutrient gradient in Great Lakes coastal wetlands. With this realistically noisy data set, differences in fit among techniques that fit gradual changes vs. ones that searched for thresholds were remarkably small, providing little evidence for superiority of one over another. However, differences in critical values among techniques were substantial, and log-transformation of total phosphorus or response endpoints altered the magnitude of critical values for all methods. Some statistical techniques require a separate step of solving for response benchmarks (issues with “zero” nutrients as a benchmark are briefly noted) whereas others identify critical stressor levels directly, but either way, professional judgment is necessary to translate analytical findings to management-relevant action levels. Since multiple statistical techniques and benchmarks provide relevant information and no transformation makes all response relationships conform to the same analysis, no cookbook recipe for stressor-response analyses can be identified. A multifaceted, synthetic approach to criteria development is warranted.