Type I Error Rates and Test Power for Some Outlier Detecting Tests

Abstract

This study is conducted to compare some outlier detection tests such as Grubbs, Dixon, Chauvenet criterion and Weisberg t-test with respect to Type I error rate and test power. For this purpose, random numbers from normal population are generated by Monte Carlo simulation technique. Results of 50,000 simulation trials showed that the Weisberg t-test gave the most reliable results with respect to the others. Chauvenet criterion has followed this test. Dixon and Grubbs tests have not displayed reliable results except for studying with small sample sizes. As a result, regardless of the sample size being studied, Weisberg t-test should be used as detecting whether there is an outlier in the data set or not.