are also not in evidence, although we have not attempted any formal statistical tests to detect clustering. We have examined plots of residuals for Iteration 1 for the wtransformed and antilog fits (not shown) and found no apparent clustering effects of. positive and negative residuals. Values of R12 for Iteration 1 for the antilog and log fits tend to be larger than for the untransformed fits (except for stratum 6 for the log fit). Iterating tends to result in smaller values of R12 for all three fits. Ideally, R12 should be zero. The residuals and observed values for stratum 4 are plotted in Figure 22 for the antilog fits, Iterations 1 and 3. These illustrate the reduction in R,* for that strata achieved by the iterating procedure. We have also computed Ro?, which is the square of the linear correlation coefficient between residuals and estimated values (Figure 23). Values of Ro? near 1 would indicate that large estimates (%.) tend to be less than the corresponding observed datum (y.), and small estimates tend to be larger than the observed datum. Figure 23 indicates values of Ro? in the range of from near zero up to about 0.20 except for stratum 6 for the antilog scale, where R22 is about 0.5. Rp* is reduced in most cases by iterating, and there appears to be little evidence to suggest the antilog or log fits are preferable to the fits in untransformed scale if Ro? is used as a criterion. CONCLUSIONS On the basis of results from three iterations, it appears that iterating on residuals can improve estimates of the true concentration surface for the Area 13 (Project 57) data set. If the results are interpreted in the original scale, fitting in units of logs and then transforming the estimated log-surface back to the original scale appears to be preferable to fitting in the untransformed scale. Alternately, the log fits could be left in log scale if interpretation in log units is desired. In making this conclusion, it is assumed that the estimated concentration surface of **!Am obtained using FIDLER is approximately the same as the true concentration surface for 239?24%Pu, Iterating on residuals tends to yield a better fitting surface to observed concentrations at sample collection points using as a criterion the size of the average absolute (mean or median) size of residuals, the standard deviation of residuals, or the proportion of the total variability in the data explained by the fit. This is true for all three scales (untransformed, log-transformed, and log). However, estimated concentrations at grid nodes (not sample collection points) are not necessarily improved This is particularly true for the untransby the iteration procedure. formed scale, where negative grid estimates are present and become even 351