Mr. Tom McCraw September 22, 1976 Page 4. It is clear that averaging plutonium concentrations will tend to reduce the apparent health risk since the peak concentrations get averaged in with the lower concentrations. This is not, however, a justification for averaging. What we need to know is what average or metric is the best indicator of future health risk to persons inhabiting the area. Guidance from resuspension and radionuclide cycling studies is needed here. Question 2: To what areas should the Pu cleanup criteria, 40 pCi/g and 400 pCi/g, be applied? This seems to be a restatement of Question 1. Again, the answer depends on how concentrations for the various size areas are related to health. If this were known and we had some idea of trends and variability over space, we would be in a better position to answer this question. Question 3: ~~~ Looking at past survey results compared with the cleanup criteria, which islands need cleanup? What levels of assurance that the criteria are met without cleanup are reasonable and attainable? A. There are a number of probability statements that can be made based on survey data. These include (1) a one-sided upper confidence limit on the true (unknown) average Pu concentration, and (2) a one-sided upper confidence limit on a percentile of the population. For this latter case, using the 95th percentile for a = .01 as an example, we could construct, e.g., an upper 100(1-a) = 99% confidence limit on the concentration level below which 95%. of the soil concentrations on the island lie. A third type of interval that appears particularly useful is a one-sided upper confidence limit on the roportion of soil concentrations that fall below the cleanup specification Tevel (this level is denoted here by L). These three kinds of limits are illustrated in an attached supplement to this letter using the ¢39-240py data collected on Janet during the 1972 Enewetak survey. We might say at this point, however, that confidence limits on average values (number 1 above) are usually computed on the assumption the data are themselves normally distributed or that the estimated mean is normally distributed. Since Pu concentrations tend to have skewed distributions similar to the lognormal, the usual procedures are sometimes modified by first transforming the data to logs, computing the limits in log scale, then transforming the limits back to the original scale. Alternatively, nonparametric or "distribu- tion-free" limits can be computed. These latter limits are valid no matter what the underlying statistical distribution, but the one-sided limits will be higher (or wider for 2-sided limits) than if a specific distribution such as the normal or lognormal is assumed. We note, however, that limits on percentiles and proportions (items 2 and 3 above) do not require any assump- | tions about the underlying statistical distribution. The several approaches mentioned above are illustrated in the Supplement.