ABSTRACT
Assessment methods fall into three broad categories. Biostatistical Methods generally apply classical statistical techniques to test null hypothesis models. These models are retrospective and cannot be used for proposed intakes or for modeling alternative operating scenarios. Predictive Biological Methods typically use intake flow characteristics and population parameters to estimate the near-term and long-term effects of the intake on a single population. Predictive methods can be used in both a retrospective and prospective manner. Community Response Methods attempt to examine the intake effects on the entire biological community (or some part thereof). While community response techniques can be used in a prospective mode, they are more commonly applied in a retrospective mode. Biostatistical Methods During the initial flurry of 316(b) activity in the mid 1970s through early 1980s, among the first methods to be applied in 316(b) assessments for the determination of AEI were biostatistical methods. These methods fell into two major groups: classical hypothesis testing using analysis of variance (ANOVA) style designs and trend analysis. ANOVA-style analyses seem to be an outgrowth of the “no prior appreciable harm” model used in 316(a) demonstrations[1] and popularized by a book on environmental sampling by Green[2]. The ideal design was often referred to as a BACI design, meaning before/after/ control/impact. A statistically significant interaction term was taken as evidence of AEI. In cases where the full BACI model could not be used, less persuasive before/after or control/impact comparisons were made. Statistically significant differences between the before and after samples or between the control and impacted samples were taken as AEI. By 1977, papers began to appear that described deficiencies in the ANOVAstyle approach[3,4] and our experienced reinforced these warnings. Problems encountered included the following:
• Failure to define the effect size to be detected • Insufficient statistical power to detect biologically meaningful impacts • Failure of untransformed data to meet requirements of normality, additivity,
independence, and heteroscedasticity • Failure of transformed data to have real-world meaning • Interpretation confounded by numerous statistically significant high-order inter-
action terms • Inability to use the model in a predictive mode (a critical problem when trying
to compare operational alternatives).
