The presence of atypical observations in a dataset is one of the causes that generate distortions in the analysis. The detection of these observations can help to evaluate the trends in the behavior of the data.In the case of multivariate data several methods have been developed that allow the detection of atypical behaviors, based on graphical methods, and others assuming a normal multivariate distribution. However, in many cases the assumption of multivariate normalcy is not fulfilled. This paper proposes a non-parametric test based on the application of Bootstrap method, using as an indicator of similarity to the distances between the representations obtained with finite series of Fourier, proposed by Andrews. The proposed method allows the detection of atypical multivariate data, combining the statistical significance of the Bootstrap test and the graphical analysis suggested by Andrews, which can be applied to data measured on an ordinal scale. The method was applied to four sets of data, finding satisfactory results in all cases.
|Published - 2018