A comparative study on univariate outlier winsorization methods in data science context

Abstract

Handling outliers is an important step in data analysis, and it can be approached through three different ways, namely; accommodation, omission, or winsorization. This article investigates the impact of four winsorization statistics (mean, median, mode, and quantiles) on parameter estimation through an extensive simulation study. Three prob- ability distributions (normal, negative binomial, and exponential) are considered, each with varying degrees of contamination. The simulation results suggest that winsoriza- tion is effective for small contamination levels and large sample sizes. Furthermore, it is recommended to winsorize outliers in symmetric distributions using any of the loca- tion parameters. However, for asymmetric distributions, the median should be employed. To illustrate these findings, a real dataset on internet usage session durations for 4,500 users, comprising over 2 million records, are fitted to the exponential distribution. The identified outliers were winsorized using the aforementioned statistics.