Outlier diagnostics in several multivariate samples

Abstract

A few diagnostics are proposed for the identification of single outliers in several multivariate samples. As in outlier detection in other models such as the linear and non-linear regression models, the proposed outlier diagnostics are found to be equivalent. The high breakdown point S-estimation robust method is used for detecting multiple outliers in several samples. The method can avoid the common masking problem in outlier detection but may tend to declare too many observations as extreme. The adding-back confirmatory analysis is used for remedying this swamping problem. Some published reference (cut-off) values are suggested for distinguishing between 'good' and outlying observations. The methods proposed are applied to real and simulated data sets. Satisfactory results are obtained.