You are here:
Improving sediment toxicity test interpretation at mega sites using novel approaches to batch normalization
Citation:
Roark, S., K. O'Neal, F. Dillon, AND Dave Mount. Improving sediment toxicity test interpretation at mega sites using novel approaches to batch normalization. SETAC North America, Sacramento, CA, November 04 - 08, 2018.
Impact/Purpose:
The logistical requirements of sediment toxicity testing programs at many contaminated sediment sites dictate that sediment samples be tested in groups, or batches, separated in time, yet the goal of most such assessments is to compare responses not just within batches, but among the entire data set. Because organism performance, especially sublethal measures like growth, biomass, or reproduction, can vary across batches, some kind of normalization of response data is generally required. Normalizing performance to the concurrent control is a common practice, but carries with it some risks for unintended consequence. Not only is it possible that the control sediment is not comparable to site sediments with low contamination, but random variation in control performance can bias data from an entire batch. This presentation explores the potential of other normalization approaches centered on normalizing to responses in site sediments representing the higher levels of organism performance, or to the median response in samples with comparatively low levels of contamination. Behavior of difference normalization practices are explored using data simulations with different underlying distributional characteristics. Results illustrate the potential for unintended bias in sediment toxicity analysis and emphasize the need to evaluate data characteristics as part of data interpretation. This information will be of interest to those charged with interpreting sediment toxicity data from contaminated sites.
Description:
Sediment toxicity tests are an important tool used in ecological risk assessments for contaminated sediment sites, but interpreting test results and defining toxicity is often a challenge. This is particularly true at mega-sites where a large testing regime necessitates performing bioassays in a sequence of batches, each evaluating a subset of the samples. Batched testing introduces variability and uncertainty that add to the challenge of interpreting results. Accounting for variability introduced by batched data collection is important. The absolute performance of test organisms is influenced by multiple factors, which contributes to variation among batches of sediment toxicity tests. These factors include the starting size, age, and health of the organisms used, conditions within each test, and random variation. For analyses that incorporate data from multiple batches of sediment exposures, the presumed goal is to parse differences in organism performance between batch-related variation and contamination-induced changes, such that the overall exposure-response gradient can be evaluated independent of batch-related variation. Variation among batches is often addressed by normalizing test sample results to control sample results, but this approach has potential for creating artifacts in the analysis. Alternative approaches include, for example, normalizing the results in each batch to a fixed percentile of the response distribution of each batch, and normalizing test results to the median (or other statistic) response in each batch associated with reference samples or with site samples determined sufficiently uncontaminated that no substantive biological response is expected. We used simulated bioassay data sets generated by resampling from log-logistic concentration response models (based on the results from real bioassays), to compare results among normalization procedures. The advantage of using simulated data was that we could separately evaluate the effectiveness of normalization approaches under specific batch variation scenarios (e.g., reduced control performance relative to test samples in some batches). The results of the simulation study illustrate the unintended results that can occur with control normalization, and the potential advantages of alternative approaches. However, there are assumptions implicit to each approach, and the most suitable approach may depend on specific characteristics of the bioassay data set.