Regional Models for Sediment Toxicity Assessment
This paper investigates the use of empirical models to predict the toxicity of sediment samples within a region to laboratory test organisms based on sediment chemistry. In earlier work, we used a large nationwide database of matching sediment chemistry and marine amphipod sediment toxicity to develop models that predict toxicity from sediment chemistry. The models estimated the probability that a sample will be toxic in two steps. First, logistic regression models were used to predict the probability that sample will be toxic based on individual chemical concentrations measured in a sample. Second, the maximum probability (PMax) of results obtained from applying the individual chemical models to a particular sample was used to predict the probability that the sample is toxic. The current paper reports on our efforts to improve model performance for regional applications. We developed four multiple chemical (PMax) models based on (1) the original published approach used to develop nationwide models (US EPA 2005); (2) a broader group of nationwide models that included additional chemicals and model variations; (3) regional models; and (4) a combination of nationwide and regional models. The models were calibrated using a data set from the NY/NJ Harbor area; model performance was compared based on classification of samples in an independent dataset from the same region. The final models developed using the calibration process substantially improved performance over the uncalibrated models developed using the nationwide dataset. The improvements were achieved by selecting the best performing individual chemical models and eliminating those that performed poorly when applied together. Although the best performing PMax model included both nationwide and region-specific models, the performance of the PMax model derived using only calibrated nationwide models was nearly as good. Because model calibration is a less intensive effort than model development, these results suggest that calibrating nationwide models to a regional dataset may be both a more efficient and effective approach for improving model performance than developing regional-specific models.