A0428
Title: Overlapping sliced inverse regression for dimension reduction
Authors: Qiang Wu - Middle Tennessee State University (United States) [presenting]
Ning Zhang - Middle Tennessee State University (United States)
Abstract: Sliced inverse regression (SIR) is a statistical tool for dimension reduction. It identifies the effective dimension reduction space, the subspace of significant factors with intrinsic lower dimensionality. We propose refined implementations of SIR algorithm by allowing slice overlapping. The new algorithms, called overlapping sliced inverse regression (OSIR), can estimate the effective dimension reduction space and determine the number of effective factors more accurately. We show that the overlapping technique codes the information of the differences (or derivatives in the population version) of the inverse regression curve, which helps to explain the superiority of OSIR. We also proved OSIR algorithms are $\sqrt n $-consistent and verified the effectiveness of OSIR algorithms by simulations and applications.
