Trimming and Winsorization are methods for dealing with the problem of outliers. This study investigated the robustness of these two procedures when there was skew in the underlying distributions. Sample data were first generated from a population with a normal distribution, and were then transformed into two skewed distributions, the lognormal and the power distributions. Trimming and Winsorization were performed on both the normal samples and transformed samples. Both Type I and Type II errors of two sample t tests and confidence intervals were investigated for evidence of robustness and power. The study found that trimmed and Winsorized means, as defined by Tukey and McLaughlin (1963) and Dixon and Tukey (1968), were not very efficient estimates of population means when the underlying distributions were skewed. While student's t test was too conservative for skewed distributions, moderately Winsorized t, as defined by Fung and Rahman (1980), was found to bring the Type I error closer to their stated levels. Furthermore, Winsorized t was found more powerful than student's t when the underlying distributions were skewed, especially when the skew was heavy, and it performed better than trimming in most cases in this study. The study suggested that these two procedures should only be used with great caution, and Winsorizing only one data point from each side of a sample was recommended with heavily skewed distributions / acase@tulane.edu
| Identifer | oai:union.ndltd.org:TULANE/oai:http://digitallibrary.tulane.edu/:tulane_24924 |
| Date | January 1998 |
| Contributors | Duan, Bin (Author), Dunlap, William P (Thesis advisor) |
| Publisher | Tulane University |
| Source Sets | Tulane University |
| Language | English |
| Detected Language | English |
| Rights | Access requires a license to the Dissertations and Theses (ProQuest) database., Copyright is in accordance with U.S. Copyright law |
Page generated in 0.0017 seconds