Pragmatic strategies for the treatment of bias in the estimation of uncertainty of measurement in analytical chemistry

摘要

In 1999, an international standard ISO 17025 made the calculation of measurement uncertainty for all analytical test results mandatory. Little guidance was available. How to account for bias in measurement uncertainty was contentious. Many analytical chemistry sectors proposed different ways of including bias in the uncertainty estimate. Six different approaches were investigated: (i) Always correcting for bias; (ii) Correcting for bias when significant; (iii) Including the bias by quadrature with the combined measurement uncertainty; (iv) Including the bias by quadrature in the expanded measurement uncertainty; (v) Adding the bias to the expanded uncertainty to give an asymmetrical confidence interval; (vi) Adding the absolute bias to the expanded uncertainty. Each of these approaches was evaluated by Monte Carlo simulation. It was found that always correcting for bias and including the uncertainty of the correction had the best outcome. If this was not possible then correcting for significant bias performed better than most other approaches. However, clarification needed to be made on what constituted the bias of a test result. Was bias to be determined within a batch run, between a batch run or between laboratories? This question was answered by looking at interlaboratory proficiency data which showed that run bias was indeed the bias of an analytical test result and therefore a correction when necessary should be made on that basis. This run bias needed to be determined across the concentration range of the analysis. Performing a regression analysis of the trueness of the analytical batch run solved this problem. The uncertainty of this regression approach was determined and applied to experimental test data. A comparison of the different approaches was then determined and it was concluded that the uncertainty based on the run bias approach gave the most realistic estimate of the uncertainty of measurement.