The influence of individual data cases in discriminant analysis, specifically in the case where an initial variable selection step precedes the analysis, is considered. An example illustrating some of the questions arising in this context, is discussed. Existing influence measures proposed by Fung (J. Amer. Statist. Assoc. 90 (1995) 952) in a non-selection context are extended to a selection context. A new more informative measure of selection influence is proposed and its properties investigated. Several examples are also presented.
Go back to Statistics seminars
Go to Department of Statistics home page