Unmasking Outliers and Leverage Points: A Confirmation

摘要: Abstract Identification of multiple outliers and leverage points is difficult because the masking effect. Recently, Rousseeuw van Zomeren suggested using high-breakdown robust estimation methods—the least median squares minimum volume ellipsoid—for unmasking these observations. These methods tend to declare too many observations as extreme, however. A stepwise analysis proposed here for confirmation detected methods. Diagnostic measures are constructed added back reduced sample. They shown graphically. The complementary use diagnostic gives satisfactory results in analyzing two data sets. One set consists often bad four good points. Four (or 10, a different cutoff) extreme other (of size 28) identified methods, but confirms only one. limitations Atkinson's confirmatory approach discusse...