The concern over outliers is old and undoubtedly dates back to the first attempt to base conclusions on a set of statistical data. The detection of outliers is important not only for univariate analysis but also for multivairate analysis. But, it is h...
The concern over outliers is old and undoubtedly dates back to the first attempt to base conclusions on a set of statistical data. The detection of outliers is important not only for univariate analysis but also for multivairate analysis. But, it is hard to derive a test procedure in the detection of multivariate outliers for the distribution of the test statistic is highly complicated. In this thesis, several detection methods for multivariate normal outliers, such as classical method by Barnett and Lewis(1978), Guttman's Bayesian technique(1973) and Varbanov's Bayesian technique(1996), are discussed and compared. The comparison is made by means of Monte Carlo simulation.
The comparison among the three method notes that, in terms of correct detection rate, Varbanov's Bayesian technique(when sequentially applied) is most efficient in detecting one or more outliers.