|
Abstract:
|
We propose a novel individualized variable selection method which performs coefficient estimation, subgroup identification and variable selection simultaneously. In contrast to traditional model selection approaches, an individualized regression model allows different individuals to have different relevant variables. The key component of the new approach is to construct a separation penalty which utilizes cross-subject information and assumes that within-group subjects share the same regression model. This allows us to borrow information from subjects within the same subgroup, and therefore improve the estimation efficiency and variable selection accuracy for each individual. Another advantage of the proposed approach is that it combines strength of homogeneity and heterogeneity in modeling and subgrouping, and therefore enhances the prediction power. We provide theoretical foundation in support of the proposed approach, and propose an effective algorithm to achieve an individualized variable selection. Simulations and an application to the HIV longitudinal data are illustrated to compare the new approach to existing penalization methods.
|
