Variable selection through adaptive MAVE

Adaptive minimum average variance estimation (MAVE) is an efficient approach for dimension reduction as it can adapt to different error distributions. In this paper, we combine the ideas of adaptive estimation and regression shrinkage, and propose the sparse adaptive MAVE (saMAVE). The saMAVE can estimate the central mean subspace and select informative covariates simultaneously, without assuming any particular model or distribution on the predictor variables. The efficacy of saMAVE is verified through both theoretical results and simulation studies.