Modified sequential fences for identifying univariate outliers

The existence of outliers in data set can bring some impacts on statistical data analysis and affect decision making. Thus, it is vital for researcher to identify the outliers. Sequential fences is a graphical method which was proposed by Schewertman and de Silva (2007). Besides its simplicity, this method is also effective in detecting multiple outliers while maintaining the approximate specific outside rate at each stage as the series on number of outlier fences. This research focuses on the modification of sequential fences to improve its efficiency. Sequential fences method is modified by replacing interquartile range with various robust scales such as semi-interquartile range, , median absolute deviation ( ) and Gini’s mean difference ( ) in order to improve outlier detection in symmetric distribution. Ultimately, the utilisation of in sequential fences seems to demonstrate a comparable accuracy in detecting the contaminated data. We have shown that GSF approach effectively reduce the masking and swamping problems in identifying the outliers. Furthermore, a new approach is proposed by considering the skewness of underlying distribution to increase efficiency of sequential fences in skewed distribution. Conclusively, based on the numerical examples and simulation study, newly proposed method has been adjusted according to the skewness of the underlying distribution of data. The results show that the new approach performed better in reducing swamping effect which is misclassifying non-contaminated observation as outlier in asymmetric distribution. Moreover, we proposed a new method with modified algorithm and methodology namely bootstrap sequential fences. The proposed method involves initial screening of data and bootstrap technique to improve the performance of sequential fences. The modified sequential fences method is found can accurately detect the outliers in positively skewed distribution. In addition, this proposed method also estimates trimmed mean and trimmed standard deviation with smaller bias and smaller root of mean squares error. Thus, proposed method proves its superiority over the existing techniques