A0372
Title: A factor screening approach for supersaturated experiments with an exponential family response via Dantzig selector 2.0
Authors: Jing-Wen Huang - Academia Sinica (Taiwan) [presenting]
Frederick Kin Hing Phoa - Academia Sinica (Taiwan)
Yu-Wei Chen - National Tsing Hua University (Taiwan)
Abstract: Dantzig selector is a powerful method for the analysis of experiments conducted in a supersaturated design. It strikes a balance between variable selection and orthogonal projection estimation. However, its underlying assumption on the normally distributed response is not always valid in real applications. The aim is to generalize the formulation of the Dantzig selector to analyze experiments with a response that follows exponential family distribution by a maximum likelihood estimation approach. It results in an approximate linear program for any fixed tuning parameters that features low computational complexity and short computational time. Moreover, we propose a binary search algorithm for tuning parameter selection. We demonstrate the performance of our proposed method via simulation studies of a supersaturated design with a logistic binary response. Our method shows good performance by comparing it with several conventional methods.
