A Comparative Study of Outliers Identification Methods in Univariate Data Set

摘要

An outlier is an observation that is inconsistent and deviate markedly from the majority of the data. The detection of outliers is a crucial part in data analysis since the contaminated observation can cause negative impacts on the data analysis. The existence of outliers can causedistortion in the data analysis and lead to erroneous conclusions. In the past literatures, there are several outliers detection methods were proposed to identify the multiple outliers in the data. In this paper, the existing outlier detection methods are discussed and reviewed. We generatecontaminated data with extreme and mild outlying observations. The performances of Tukey’s boxplot, Hampel’s test, modified Z-score method, sequential fences and extreme standardized deviate test are applied on a simulated data set in order to investigate extensively theirsensitivity towards outliers and the time consuming of each technique in detecting the outliers. The results suggest that the Tukey’s boxplot method offers substantially better performance and sequential fences procedure shows good results under certain confidence level.