Multinomial Regression with Variable Selection and Outcome Merging Penalty

• Activity Number: 570
• Type: Contributed
• Date/Time: Wednesday, August 1, 2012 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Sheng-Mao Chang*+
• Companies: National Cheng Kung University
• Keywords: LASSO ; latent class analysis ; multinomial regression

Abstract

In this work, we apply the lasso regularization method to the multinomial regression. Multinomial regression is a useful tool to discover the association between nominal outcomes and their covariates. In general, different outcomes have different associations with these covariates. Sometimes, two or more outcomes are not distinguishable in the sense that they share common regression coefficients and then it is reasonable to merge these outcomes into one. To this end, two penalty terms are introduced to select important covariates and to merge indistinguishable outcomes. This approach can be an alternative of latent class regression with certain classification information. Contrasting to the latent class method, an attractive advantage of the proposed approach is that it estimates not only the regression coefficients but also the number of the latent classes.