Detecting bivariate outliers on the basis of normalizing transformations for non-Gaussian data

Abstract

The statistical technique for detecting outliers in bivariate non-Gaussian data on the basis of normalizing transformations, prediction ellipse and a test statistic (TS) for the Mahalanobis squared distance (MSD), which has an approximate F distribution, is proposed. Application of the technique is considered for detecting outliers in two bivariate non-Gaussian data sets: the first, actual effort (hours) and size (adjusted function points) from 145 maintenance and development projects, the second, effort (hours) and mass (tonnes) of designed the section of the ship from 188 designs of sections.