1.21
Quantifier et rejeter les valeurs aberrantes : le test de Grubbs
Parfois, un ensemble de données peut contenir une observation numérique enregistrée qui s’écarte considérablement du reste des données. En supposant q…
Le test de Grubbs, comme le test Q de Dixon, est un test statistique permettant d'identifier les valeurs aberrantes dans les données ayant une distribution normale. Ici, le nombre d’observations doit être égal ou supérieur à sept.
Pour un ensemble de données donné, ce test consiste à calculer la statistique de Grubbs G, qui est le rapport de la différence absolue entre la valeur douteuse et la valeur moyenne et l'écart-type de l'échantillon.
La valeur G calculée est ensuite comparée à la valeur critique tabalée de G pour un niveau de confiance donné et le nombre d’observations.
Si la valeur G calculée dépasse la valeur critique de G, l’observation douteuse est considérée comme une valeur aberrante des données et rejetée.
Alternativement, si la valeur G calculée est inférieure à la valeur critique de G, les données conservent l’observation douteuse.
View the full transcript and gain access to JoVE Core videos
Q1: What is the Grubbs test used for in analytical chemistry?
The Grubbs test is a statistical method for identifying outliers in normally distributed data sets. It determines whether a questionable observation that greatly deviates from the rest of the data is truly an outlier or should be retained. This test requires a minimum of seven observations and compares a calculated statistic against a critical value to make the rejection decision.
Q2: How do you calculate the Grubbs statistic G?
The Grubbs statistic G is calculated as the ratio of the absolute difference between the questionable value and the mean to the standard deviation of the sample. First, subtract the mean from the suspected outlier and take the absolute value. Then divide this difference by the sample's standard deviation. This G value is then compared with the tabulated critical value to determine if rejection is warranted.
Q3: When should you reject an observation using the Grubbs test?
An observation is rejected as an outlier when the calculated G value exceeds the critical value of G for your specified confidence level and number of observations. Conversely, if the calculated G value is smaller than the critical G value, the questionable observation is retained in the data set. The critical value serves as the threshold for making this rejection decision.
Q4: What are the minimum requirements for applying the Grubbs test?
The Grubbs test requires that your data set contains at least seven observations and follows a normal distribution. The test assumes the data is normally distributed before you calculate the Grubbs statistic. Meeting these prerequisites ensures the statistical validity of your outlier detection and rejection analysis.
Q5: How does the Grubbs test differ from Dixon's Q-test?
Both the Grubbs test and Dixon's Q-test are statistical methods for identifying outliers in normally distributed data. While the transcript mentions they serve similar purposes, the Grubbs test specifically calculates a ratio of the absolute difference between the questionable value and mean to the standard deviation, then compares it against tabulated critical values for a given confidence level.
Q6: What role does confidence level play in the Grubbs test?
The confidence level determines which critical value of G you use for comparison with your calculated G statistic. Different confidence levels yield different critical values from statistical tables. Your choice of confidence level affects whether a questionable observation is rejected or retained, making it a crucial parameter in the outlier detection decision.
Q7: Why is the Grubbs test limited to normally distributed data?
The Grubbs test assumes data follows a normal distribution because the statistical theory underlying the test relies on this assumption. The calculation of the Grubbs statistic and the critical values in reference tables are derived based on normal distribution properties. Applying the test to non-normally distributed data would compromise the validity of your outlier detection results.