Factor Profiling for Ultra High Dimensional Variable Selection

2010; RELX Group (Netherlands); Linguagem: Inglês

10.2139/ssrn.1613452

ISSN

1556-5068

Autores

Hansheng Wang,

Tópico(s)

Advanced Statistical Methods and Models

Resumo

We propose here a novel method of factor profiling (FP) for ultra high dimensional variable selection. The new method assumes that the correlation structure of the high dimensional data can be well represented by a set of low-dimensional latent factors (Fan et al., 2008). The latent factors can then be estimated consistently by eigenvalue-eigenvector decomposition. They should be profiled out subsequently from both the response and predictors. Such an operation is referred to as FP. Obviously, FP produces uncorrelated predictors. Thereafter, the method of sure independent screening (Fan and Lv, 2008, SIS) can be applied immediately. This leads to profiled independent screening (PIS). PIS is shown to be selection consistent, even if the predictor dimension is substantially larger than the sample size. To further improve PIS, a novel method of profiled sequential screening (PSS) is proposed. PSS shares similar strength as forward regression (Wang, 2009a) but is computationally even simpler. Numerical studies are presented to corroborate our theoretical findings.

Referência(s)
Altmetric
PlumX