Methods for Evaluating the Reference Value in Laboratory Intercomparisons of Dimensional Measurements
Jack A. Stone Jr.
A number of methods have been proposed to evaluate the reference value for intercomparisons of laboratory measurements. Methods for establishing the reference value include the arithmetic mean, weighted mean (with weights proportional to the reciprocal of the squared uncertainty), median, and total median. In addition, it has been suggested that it might be possible to modify the weighted mean, using iterative approaches to automatically eliminate outliers or to modify the weights in light of the results of the intercomparison. No single one of the analysis methods is best for all circumstances, nor can the efficiency of any method be determined without making assumptions about the underlying nature of the intercomparison. (How well do the participants estimate their uncertainties? What is the underlying distribution of errors, including outliers? Are the errors correlated between one laboratory and the next?) Although there is considerable divergence of opinion as to what constitute realistic assumptions, completed international comparisons can begin to provide at least rough guidance for constructing models. In this paper I try to construct models that are consistent with what we have learned thus far for CCL (Consultative Committee for Length) key comparisons in the field of dimensional metrology. Based on such models, I have explored various methods for establishing a reference value, to determine which methods are likely to produce a reference value with a low uncertainty. As would be expected, there is no single method that is always superior; results depend on both the underlying assumptions and on the spread and distribution of claimed uncertainties of the participating laboratories.
Proceedings of SPIE
July 31, 2005
San Diego, CA
Recent Developments in Traceable Dimensional Measurements III
comparison, evaluation, key comparison reference value