Grubbs Outlier Test
 |
|
Description: This script performs a utility function. It detects an outlier in univariate data using the method of Grubbs (1,2). It is used to check one data column at a time. It is used iteratively, that is, if you find an outlier, delete it and repeat the procedure. The script recognizes excluded rows so if you detect an outlier, simply select this row and exclude it, then run the script again to find the next one. The report includes a statement if the outlier is or is not detected at the significance level indicated. It also gives a p-value for the observed G statistic.
Note the test should not be used with samples of 6 or less. This recommendation is not enforced by the script, however.
Note this test assumes a single Gaussian distribution and no particular kind of contamination.
(1) NIST/SEMATECH Engineering Statistics Handbook, Section 1.3.5.17, http://www.itl.nist.gov/div898/handbook/eda/section3/eda35h.htm.
(2) V. Barnett and T. Lewis, "Outliers in Statistical Data, 3rd Edition," Wiley (1994).
Instructions: First open the data table below, then open and run the script. Select the weight column and click Y button. Leave the significance level (alpha) at 0.05 for now. Click OK. Check the normal quantile plot for evidence of non-normal distribution and outliers. The new report appears at the bottom of the window, below the closed Quantiles and Moments reports.
Go back to the data table and change the first weight (Donna) to 108 and repeat the analysis. This value is just outside the critical interval and so it is flagged.
Requirements: Fitness.jmp (included)
SAS Campus Drive, Building S, Cary, NC, 27513 • Phone: 1.919.677.8000© 2008 SAS Institute, Cary, NC, USA. All Rights Reserved. Terms of Use • Privacy Policy