Diskussionspapiere // Friedrich-Alexander-Universität Erlangen-Nürnberg, Lehrstuhl für Statistik und Ökonometrie 09/1996
Statistical analysis is often disturbed by objects which are extremely different from the rest of the data. Those outliers can be due to different causes. Therefore it is always recommended to examine them separately. Outliers in one or two-dimensional cases are easily recognized in a frequency distribution. In multidimensional data they can be identified by sub-ordering. It is proposed to do this by calculating pairwise distances. The necessary standardization of the variables can be done by using the sum of all pairwise distances. Proceeding this way possible outliers can easily be identified in tabular as well as in graphical form. It can also be demonstrated which dimension, that is which variables are contributing to the outlier status. As it is quite simple to remove any object or variable, one can see what is happening to the rest of the data without those extreme values.