All contamineint concentration data to be used in a risk assessment 

 should undergo a thorough QA review by a qualified chemist inde- 

 pendent of the laboratory that analyzed the samples. In some cases, the 

 analytical laboratory may provide a QA review that is simply checked 

 by an independent chemist. The purpose of the QA review is to 

 evaluate the data relative to data quahty objectives (e.g., precision and 

 accuracy) and performance limits estabhshed in the QA plan. In many 

 cases, qualifiers are necessary for selected data values. These qualifiers 

 should be added to the database. A summary of data limitations should 

 always be included in the risk characterization (see below, Risk Char- 

 acterization). The EPA Office of Acid Deposition, Environmental 

 Monitoring, and Quality Assurance is developing guidelines for quality 

 assurance of chemical data to support exposure assessments. 



Statistical analyses of data will depend on specific study objectives. For 

 each species, statistical summaries of tissue concentration data should 

 include sample size, estimates of arithmetic mean concentration, 

 range, and a measure of variance (standard error or 95 percent con- 

 fidence limits). Geometric mean concentrations are appropriate 

 measures of central tendency when only estimates of tissue burden of 

 contaminants or exposure dose are desired. However, arithmetic 

 means are needed to compare exposure estimates with RfDs and to 

 calculate health risk for chronic effects because long-term consump- 

 tion is an averaging process. Mean tissue concentrations and variances 

 may be calculated for mixed-species diets if data are available on the 

 proportion of each species m the diet. 



The one-way ANOVA model discussed earlier or multifactor ANO VA 

 models are appropriate for testing for differences in mean contaminant 

 concentrations among species, among sampling stations, or among 

 time periods (Schmitt 1981; also see Tetra Tech 1986b,d). For small 

 sample sizes and data that do not satisfy the assumptions of ANOVA, 

 nonparametric tests such as the Wilcoxon rank sum test for two treat- 

 ments or the Kruskal-Wallis test for multiple comparisons are recom- 

 mended. These tests have the added advantage of being relatively 

 insensitive to a few missing data points or undetected observations 

 (Gilbert 1987). Long-term data sets may be tested for trends by time 

 series analysis (for reviews, see Montgomery and Reckhow 1984 and 

 Gilbert 1987). Examples of trend analysis for chemical contaminants 

 in fish are provided by Brown et al. (1985b) for PCBs in striped bass 

 of the Hudson River and by DeVault et al. (1986) for PCBs and DDT 

 in lake trout from the upper Great Lakes. 



Data on concentrations of contaminants of concern in tissue samples 

 will often contain observations below detection limits. Means and 

 variances for tissue concentrations should be calculated twice: once 

 using detection limits for undetected observations and once using for 

 undetected observations. Although alternative approaches are pos- 

 sible (e.g., using one-half the detection Hmit), the approach recom- 

 mended here yields more accurate, complete results by quantifying the 

 range of the estimated values. According to the EPA Exposure Assess- 

 ment Group, calculations of plausible-upper-limit risk estimates based 

 on detection limits should generally be avoided. However, risk es- 

 timates based on detection limits may occasionally be useful to indicate 

 that particular chemicals, species, or geographic locations are not 



Statistical Treatment of Data 



51 



