Statistics Glossary → Grubbs' and generalized ESD tests
Grubbs' and generalized ESD testsconcept
Tests of whether the observation farthest from the mean (Grubbs) or up to r observations (generalized ESD) are outliers, in data assumed to be normally distributed.
Grubbs' test uses G = max|xᵢ − x̄| / s: it measures how many standard deviations the most distant observation lies from the mean and compares this with the largest deviation expected in a normal sample of size n. It is designed to test a single outlier. Applying it repeatedly is problematic: several outliers can hide each other by inflating the standard deviation (masking).
Rosner's generalized ESD (extreme studentized deviate) test, for a pre-specified maximum of r outliers, removes the most extreme observation step by step, recomputes the statistic each time and compares it with adjusted critical values, which makes it more resistant to masking. Both tests assume that the data are normal apart from the outliers; with skewed distributions they can flag genuine extreme values as outliers. Finding an observation statistically outlying is not reason enough to delete it; its cause should be investigated.
An everyday example
A laboratory measures the pH of the same sample 12 times; the values lie between 7.01 and 7.08, except one reading of 7.45. With a mean of 7.08 and a standard deviation of 0.12, G ≈ (7.45 − 7.08) / 0.12 ≈ 3.1, which exceeds the critical value of about 2.4 for 12 observations. The lab notebook shows that the electrode was not calibrated before that reading.
In MerQur
In the Anomaly Detection analysis, as two of the ten methods offered: Grubbs' test and Rosner's ESD; an anomaly reason and a flag column are given for each observation.