Identifying multivariate outliers - a medical example

Abstract

Many data sets, especially medical data, consist of a two-dimensional table Xnxp containing p variables measured for every of n individuals. We are concerned with values of p=9 traits, such as Age, Height and other spirometric variables like RV, VC, VC%, FEV1, FEF,... recorded for n=125 patients. Such table can be interpreted as a cloud of n points in the p-dimensional Euclidean space Rp. The analysed data contain outliers both in size and structure. Especially the last type could not be detected when considering each variable individually. We demonstrate the usefulness of modern visualization methods for multivariate data, as grand tour with a count plot (Bartkowiak and Szustalewicz, 1997) which finds a set of points suspected to be outliers, and then, the complete linkage method (based on angular distances) and parallel coordinate plot - which additionally confirm the obtained results.