Home | Search | Help  
Home Page Università di Genova

Seminar Details


Date 7-12-2007
Time 15:00
Room/Location Sala conferenze-DISI-3 piano
Title Two consistent approaches to variable selection in a constrained and unconstrained framework
Speaker Dott.Alessio Farcomeni
Affiliation Universita' La Sapienza, Roma.
Link http://www.uniroma1.it/
Abstract Methods based on maximization of an L1 penalized likelihood, like the LASSO, are widely used due to their properties of simoultaneous shrinkage and selection and their relatively low computational burden. On the other hand, LASSO based methods may not be consistent. When a true model is believed to exist, it may be desireable to use methods that are guaranteed to choose it, at least asymptotically. We illustrate two consistent methods which still have low computational burden. First we propose to fit the model in a two-step fashion, with a multiple testing approach. Secondly, we illustrate the Bayesian approach to simoultaneous shrinkage and selection, which can usually be implemented with a fast and efficient Gibbs sampler. We show that the median posterior model is consistent under very wide conditions, and propose a reparameterization of the usual stochastic search method to accomodate constraints in variable selection.
Back to Seminars