A procedure for the detection of multivariate outliers

Single case diagnostics are susceptible to a masking effect. This has led to the development of methods for detecting of multiple multivariate outliers. The available methods work well but may not be able to always detect outliers in data with contamination fraction greater than 35%, as reported by Rocke and Woodruff (1996, J. Am. Statist. Assoc. 91, 1047-1061). In this paper we propose a new method for detection of outliers which is very resistant to such high contamination of data with outliers. The simulation results indicate that, while maintaining the nominal level, the proposed method is never worse and detects outliers better than the Rocke and Woodruff method for data highly contaminated (35-45%) with outliers. Improved performance was also noted for data with smaller contamination fraction (15-20%) when outliers were situated closer to the 'good' data. Several data sets are used to illustrate the proposed procedure.

Duke Scholars

Author Andrzej Stanislaw Kosinski Biostatistics & Bioinformatics, Division of Biostatistics