A0848
Title: Good subsets approach to variable selection
Authors: Neill Smit - North-West University (South Africa) [presenting]
Riaan de Jongh - North-West University (South Africa)
Hennie Venter - North-West University (South Africa)
Abstract: A recently proposed variable selection procedure is discussed and compared to other well-known variable selection procedures. The lambda-good variable selection procedure is based on selecting good subsets at a specified margin lambda. A subset is said to be good at margin lambda if the associated criterion of fit improves in relative terms if any variable is added, and the criterion of fit deteriorates in relative terms by at least lambda if any variable is dropped. The performance of the lambda-good procedure in terms of variable selection accuracy and computational efficiency will be demonstrated and compared in a simulation study to existing procedures, such as best subsets, forward stepwise selection and the lasso.
