Métodos para detecção de Outliers multivariados: Via uso dos estimadores robustos

Abstract

In the application of the multivariate analysis, it is necessary to follow some procedures in order

not to obtain an erroneous relationship between the phenomenon of interest and the other varia- bles, that is, the model needs to be well adjusted to the characteristics of the phenomenon under

study. The detection of outliers is an important method to be applied in statistical analyses,

because a single outlier can cause changes in parameter estimates, also interfere with norma- lity and correlation tests between variables, in addition to alter the results of any other inference

procedure. Therefore, the objective of this work is to present and compare some methods for de- tecting outliers in multivariate data. The minimum volume ellipsoid (MVE), minimum volume

covariance (MCD), orthogonalized Gnanadesikan and Kettenring (OGK) methods, principal components for detection of outliers (PCOut) and Comedian were compared. To perform the

comparisons, a series of simulations was used, predicting different situations using the conta- minated normal distribution. Comparisons were evaluated through the success rate (TS), which

indicates the percentage of outliers that the methods correctly identified, and the false detec- tion rate (TFD), which indicates the percentage of observations that are not outliers, but were

identified as outliers. It is concluded that the ideal is to use at least two methods to detect ou- tliers, since pointing out the only method as the best is a difficult task. However, the PCOut

and Comedian methods obtained the best TS in most of the simulated scenarios. The comedian method obtained the best TFD.