Contributed paper list
(Tuesday 15th, 13:45-15:15) In session C7C
A comparison of outlier labeling criteria in univariate measurements
Menus Nkurunziza, Lea Vermeire
Presenter Menus Nkurunziza
For the detection of potential outliers in univariate measurements, undergraduate statistics courses often refer to the boxplot. In the workfield, various other sector-linked criteria for outliers are also popular, e.g. Chauvenet’s criterion in engineering. We compare statistical properties of five current criteria – the 3-sigma rule, the Z-score, Chauvenet’s criterion, the M-score or median criterion, and the boxplot or Tukey’s criterion. In particular, in case of a normal population, a joint structure of the five criteria is detecte,d and large sample asymptotic properties of their non-outlier intervals are derived. Pointing at these results should help students to match the statistics course and the lab practice during their education or in their future professional environment. Next, for mathematical statistics students, proving these results may be an instructive activity.