Approaches to handling discordant observations: an appraisal

Performances of several methods currently used for detection of discordant observations are reviewed in terms of a data set pertaining to absolute measurements of gravity acceleration, showing some peculiar features. Along with currently used methods a criterion based upon distribution of extremes is also relied upon to provide references; a modification of a simple, broadly used method is mentioned, improving performances while retaining inherent ease of use. Identification of distributions underlying experimental data may entail a substantial uncertainty component, particularly when sample size is small, and no mechanistic models are available. A pragmatic approach is described, providing to a first approximation estimation of overall uncertainty, inclusive of components pertaining to both estimation of parameters, and identification of distribution shape.