A Method for Oultier Detection in Grouped data

The method proposed by Hadi (1994) for multiple outlier detection in a single group of multivariate data is adapted to the multiple cluster setting. The idea is to replace, in Hadi’s algorithm, the Gaussian distribution and the Mahalanobis distance with the K-component normal mixture model (with K > 1) and a coherent measure of discrepancy from a mixture distribution, respectively. The performance of the proposed procedure is illustrated on a real data set and compared, through a simulation study, with the method proposed by Caroni and Billor (2007) for detecting multiple outliers in grouped multivariate data.