The Analyst · 2001 · 121 citations · 2 references
EngineeringCertified Reference MaterialsConformance TestingExperimental TestingAnalytical ChemistryTestabilityStatisticsMaterials ScienceReliabilityMethod ValidationHomogenization TheoryNondestructive TestingChemometricsFormal TestMaterials CharacterizationNew TestData HeterogeneityAnalytical MethodMaterials Testing
Certified reference materials and proficiency‑testing distributions must be “sufficiently homogeneous,” meaning the variance in the mean composition of distributed portions must be negligible relative to the analytical result variance. The study proposes a new statistical test for sufficient homogeneity to address shortcomings of existing methods. The authors develop a simple method based on duplicated portion analyses and ANOVA to assess homogeneity. They find that current tests can misclassify materials—over‑rejecting homogeneous samples when the analytical method is very precise, under‑detecting heterogeneity when it is imprecise, and that the Harmonised Protocol Procedure tends to reject satisfactory material.
Certified reference materials and materials distributed in proficiency testing need to be 'sufficiently homogeneous', that is, the variance in the mean composition of the distributed portions of the material must be negligibly small in relation to the variance of the analytical result produced when the material is in normal use. The requirement for sufficient homogeneity suggests the use of a formal test. Such tests as have been formulated rely on the duplicated analysis of the material from a number of portions, followed by analysis of variance. However, the outcome is not straightforward. If the analytical method used is very precise, then an undue proportion of the materials will be found to be significantly heterogeneous. If it is too imprecise, the test may be unable to detect heterogeneity. Moreover, the Harmonised Protocol Procedure (M. Thompson and R. Wood, Pure Appl. Chem., 1993, 65, 2123) seems to be unduly prone to the rejection of material that is in fact satisfactory. We present a simple new statistical approach that overcomes some of these problems.
2
A confidence interval for variance components
J. S. Williams · Biometrika · 1962 · 84 citations