A Generalized Outlier Method with an Automated Test Protocol Applied to Proficiency Testing Schemes: Utilizing Monte Carlo Simulation for Generating Critical Values for the Extended Grubbs Statistic.

BACKGROUND: A generalized method for addressing outliers in Proficiency Testing (PT) schemes is presented, along with examples illustrating their impact on the analysis used for payment in the grain industry. OBJECTIVE: Enhance the accuracy and robustness of the assigned values in the newly proposed automated testing protocol. Additionally, provide added value to participants in PT schemes by enhancing the significance of outlier detection, enabling further investigation of the underlying causes. METHODS: The Grubbs outlier detection method has been expanded to include 35 distinct tests for identifying single, double, and up to multiple-10 outliers. Monte Carlo simulations have been used to determine the critical values for each test, which are suitable for datasets containing up to 1000 observations. This research specifically focuses on the one-sided significance level of 1.25%, but six other levels are included. The application of this method is particularly relevant for analyzing PT scheme data within the grain industry. RESULTS: The study discusses the challenges posed by extreme outlier data that may have identical values due to rounding practices. Additionally, another example from the world's largest grain network in France shows a multiple-5 outlier in a large data set of over 700 observations. CONCLUSION: This extended Grubbs test provides a more accurate determination of mean values and facilitates the identification of outlying observations, which may reveal significant insights when confirmed to be statistically distinct. HIGHLIGHTS: The importance of precise testing using the new protocol is underscored by the observation that even a minimal shift of 0.1% in protein content in barley could lead to financial implications amounting to millions of US dollars. Notably, a minor shift of 0.07% was observed when applying the quadruple outlier test that excluded the four highest observations from the PT scheme dataset.