Outliers Detection for Pareto Distributed Data

This study aims to examine the presence of outliers in the upper tail of Malaysian income distribution under the assumption that the data follow Pareto model. For this purpose, three types of boxplot: standard boxplot, adjusted boxplot and generalized boxplot are considered. The performance of these boxplots is determined by a simulation study. In this study, the data were simulated from Pareto distribution, P(1, alpha = 2, 3, 4), then the simulated data were contaminated by replacing a proportion epsilon (3%, 5%, 10%) of randomly selected data. It is found that the generalized boxplot gives higher power value compared to the standard and adjusted boxplots. Therefore, the generalized boxplot was used for determining the presence of outliers in the upper tail of income distribution, while the threshold for Pareto tail modelling was determined by using Van Kerm's formula. The results showed that 0.4%, 0.4%, 0.9% and 1.2% outliers were detected by the generalized boxplot in the household income data that exceeded the threshold for the years of 2007, 2009, 2012 and 2014.