Comparison of Multiple Linear Regression and Biotic Ligand Models to Predict the Toxicity of Nickel to Aquatic Freshwater Organisms
Kelly Croteau, Adam C. Ryan, Robert C. Santore, David K. DeForest, Christian E. Schlekat, Elizabeth Middleton, Emily Garman
Abstract
Toxicity-modifying factors can be modeled either empirically with linear regression models or mechanistically, such as with the biotic ligand model (BLM). The primary factors affecting the toxicity of nickel to aquatic organisms are hardness, dissolved organic carbon (DOC), and pH. Interactions between these terms were also considered. The present study develops multiple linear regressions (MLRs) with stepwise regression for 5 organisms in acute exposures, 4 organisms in chronic exposures, and pooled models for acute, chronic, and all data and compares the performance of the Pooled All MLR model to the performance of the BLM. Independent validation data were used for evaluating model performance, which for pooled models included data for organisms and endpoints not present in the calibration data set. Hardness and DOC were most often selected as the explanatory variables in the MLR models. An attempt was also made at evaluating the uncertainty of the predictions for each model; predictions that showed the most error tended to show the highest levels of uncertainty as well. The performances of the 2 models were largely equal, with differences becoming more apparent when looking at the performance within subsets of the data. Environ Toxicol Chem 2021;40:2189-2205. © 2021 SETAC.