Statistics Seminar Series: Efficient and Stable Box-Cox Regression for High-Dimensional Data Analysis

Room 151. Efficient and Stable Box-Cox Regression for High-Dimensional Data Analysis Please join the Statistics Department for a talk featuring Hui Zou, professor in the Department of Applied Mathematics and Statistics at Johns Hopkins University. The Box-Cox transformation paper ranks among the five most-cited articles in the Journal of the Royal Statistical Society, Series B. Bickel and Doksum (1981) revisited this classical model through analysis of the profile likelihood estimator and concluded that "the performance of all Box-Cox type procedures is unstable." We argue, however, that the scientific modeling philosophy of Box and Cox remains valuable in modern data analysis. To this end, we develop a novel composite likelihood framework for estimation and inference in a high-dimensional nonparametric Box-Cox (NBC) model. The composite likelihood approach directly resolves the instability concern of Bickel and Doksum (1981). Furthermore, NBC yields substantially improved prediction accuracy over standard regression, while preserving its core advantages: computational efficiency and model transparency. Hui Zou's research interests include high-dimensional statistics, machine learning and statistical optimization. Zou is a recipient of the ICSA Pao-Lu Hsu award and ISI Founders of Statistics prize. He is an elected Fellow of IMS, ASA and AAAS.