The following HTML text is provided to enhance online
readability. Many aspects of typography translate only awkwardly to HTML.
Please use the page image
as the authoritative form to ensure accuracy.
be representative of in situ physicochemical conditions because some parameters (such as pH level) can vary naturally over daily cycles, and others (such as conductivity) may change strongly in response to waste-water discharges. These two issues were explored by using Ceriodaphnia tests to evaluate water-quality conditions in upper East Fork Poplar Creek, where TRC was suspected of causing or contributing to fish kills. Logistic regression was used to relate TRC data to toxicity test outcomes (Stewart et al., 1996).
We first analyzed the chemical data (daily measurements of pH, conductivity, alkalinity, hardness, and TRC) for 169 site and test-period combinations (4 sites were tested over a 50-month period). For each water-quality factor, we computed a 7-day mean and an estimate of daily variability, referred to as semirange. Semirange was defined as a parameter's 7-day maximum (transformed) value minus the 7-day mean. For toxicity assessments, one advantage of semirange is that it quantifies excursions above the mean but ignores excursions below the mean. (Toxicologically, pollutant concentrations above the mean are likely to be more significant than those below.) We then used stepwise logistic regression to explore relationships between the 7-day mean and 7-day semirange values for the water-quality factors and Ceriodaphnia mortality. Both the proportion of animals dying in each test and the pass-or-fail outcomes (using 60 percent survival as the pass-or-fail criterion [see Table 1]) were assessed.
The results of these analyses showed that 7-day mean TRC concentration and TRC semirange both strongly affected Ceriodaphnia mortality (p < 0.0001 for each factor). With these two factors included, the logistic regression model correctly predicted the outcome (mortality or survival, expressed as a proportion of the animals tested in each test) in 89.3 percent of the cases. The model's false positive rate (when the model predicted mortality, but no mortality occurred) was 20 percent, and the model's false negative rate (no mortality was predicted by the model but the animal died) was 7 percent. Distilling a test's outcome to a passor-fail status using the criterion of 60 percent survival was a satisfactory simplification: Both TRC mean and semirange values were significant as explanatory factors (p < 0.0001 in each case, with 91.7 percent of the cases being predicted correctly by the model), and the model's false positive rate and false negative rates were low (15.2 and 5.7 percent, respectively).
Figure 3 is a schematic showing the generalized flow for Ceriodaphnia toxicity test data used in the statistical analysis methods described in this paper. Various data-checking steps cited for use in the effluent data flow path (e.g., inspection of variance for homogeneity [Figure 2]) are also appropriate when analyzing ambient toxicity test data, but these are not shown in Figure 3 for convenience.
Diagnostic Testing and Ambient Toxicity Monitoring
The logistic regression study described above also demonstrated that diagnostic or ''experimental" toxicity testing should be integrated into any ambient