February 4, 1999, 16.15 - ca. 17.30
Hauptgebäude der Universität, Hörsaal E 18
We first discuss the robustness aspects related to the choice of a model and review some ideas available in the literature. In particular we discuss robust versions of classical model selection procedures like Mallows's Cp and Akaike's Criterion.
In the second part of the talk we provide a robust procedure for model selection in regression using cross-validation methods and estimators which have optimal bounded influence for prediction. The results are contrasted with those obtained by a least squares analysis demonstrating a substantial improvement in choosing the correct model(s) in the presence of deviations with little loss of efficiency at the normal model.