Abstract

Outlier is an object that has unique characteristics compared with other objects. Detection of outlier needs to be performed in order to avoid errors in decision-making related to data. Another reason for the detection of outlier is to determine the cause and the meaning of the difference from the outliers. Two methods for detection of outlier are Minimum Covariance Determinant (MCD) and Fast Minimum Covariance Determinant (FMCD). Unfortunately, MCD and FMCD require a longer iteration thus it is difficult to detect outlier in large data. In this work, we introduce alternative methods to detect outliers i.e. direct and indirect biplot analyses and direct and indirect Euclidean distances and then we determine the effectiveness of methods in detecting outlier using Procrustes measure. The larger the size of Procrustes measure, the better the method for detecting the outliers. There are six data used in the simulation analysis, i.e. the generated data with the characteristics of one group data, one group data with the outlier, one group data with the top and bottom outliers, two-group data, three-group data. Inferred data criteria based on simulation analysis results that MCD, FMCD, indirect biplot and indirect Euclidean distance can’t detect outlier of grouped data. Applicative analysis is the detection of outliers in the welfare data of the Indonesian people based on seven indicators. Papua and DKI Jakarta provinces are concluded as outliers based on all methods. Further analysis reveals that direct Euclidean, indirect Euclidean, and indirect biplot are the best methods. However, direct Euclidean is the simplest method.

Authors and Affiliations