Garrote Methods for Generalized Linear Models

Keywords: Prediction, shrinkage, stability, subset selection

Abstract: Hand crafted subset selection procedures are often used for generalized linear models. However, subset selection is known to be unstable and relatively inaccurate.

Recently, more stable methods have been proposed for selecting subsets and shrinking in the regression context. One such is the nonnegative garrote (Breiman, 1995). We extend the garrote to generalized linear models and introduce a new method for computing garrote solutions.