On the adjustment of inconsistent data using the Birge ratio

Abstract

The Birge ratio is applied in metrology to enlarge quoted uncertainties when combining inconsistent measurement results on the same measurand. We discuss the statistical model underlying such a procedure and argue that the resulting uncertainty associated with the adjusted value is underrated. We provide a simple modification of this uncertainty on the basis of an objective Bayesian inference. While the proposed uncertainty approaches that obtained by the conventional procedure for a large number n of combined measurement results, differences are significant for small n. For example, for n = 4 we get an increase of 73% in the standard uncertainty associated with the adjusted value, and for n = 10 the increase is still 13%. We derive the posterior distribution for the adjusted value in closed form, including a 95% credible interval. In addition, we show that our results do not only hold when the distribution of the measurement results is assumed to be Gaussian, but for a whole family of (elliptically contoured) location-scale distributions. We illustrate the modified Birge method by its application to data from the 2002 adjustment of the Newtonian constant of gravitation.