An iterative method for improved estimation of the mean of peer-group distributions in proficiency testing

In proficiency testing (PT), the peer-group mean is conventionally computed after twice removing values exceeding the mean +/- 3 SD. However, this adjustment fails if there are many outliers. In this study an iterative method was evaluated as a more robust way to estimate the means. The methodology repeatedly removes a proportion of the population (usually those exceeding the mean +/- 1.6 SD), assuming the presence of a Gaussian distribution in the central portion, and reinflates the SD to compensate for the trimming. A computer simulation revealed that the estimated mean of a known Gaussian distribution was less affected by a subpopulation that overlaps the main population than was the conventional method. When the overlapping portions were removed, the iterative method predicted the true mean correctly. The method was applied to external PT results for 44 analytes. Although most peer-group distributions were clearly non-Gaussian, the segment included by the predicted mean +/- 1.6 SD was regarded as Gaussian in 85.9% by the new method and 73.4% by the conventional method. The proposed methodology appears to be an improved way of estimating peer-group means.