An adaptable generalization of Hotelling's \(T^2\) test in high dimension (Q2196248): Difference between revisions

There is an investigation on the two-sample testing problem for high dimensional means in connection with the Hotelling's \(T^2\) approach. A test statistic based upon the regularized Hotelling's \(T^2\), RTH, introduced by \textit{L. S. Chen} et al. [J. Am. Stat. Assoc. 106, No. 496, 1345--1360 (2011; Zbl 1234.62082)] for the one-sample case, is proposed and extensively discussed. The authors provide a Bayesian framework to analyze the power of the RTH and construct a new composite test by combining the RTH statistics corresponding to a set of optimally chosen regularization parameters. The new composite testing procedure will be named ``adaptable RTH'', shortly ARTH. The weak convergence of the considered stochastic process to a Gaussian limit is proved. The asymptotic behavior of the test by relaxing the assumption of Gaussian to sub-Gaussian is investigated. The regularized Hotelling's \(T^2\), RTH, test is introduced and largely discussed in the second section of the article while the adaptable ARTH is presented in the third section. The calibration of type I error probability and extension to a general class of sub-Gaussian distributions are presented in the fourth and fifth section. Simulation results are shown in the sixth section and a practical application to a breast cancer data set in the seventh section. In the eighth section, we find a short discussion on obtained results. The ninth section is devoted to the proofs of main results and is followed by a short appendix containing some auxiliary results. One reports that additional simulation results and detailed proofs of the main results are contained in a supplementary material in \url{doi:10.1214/19-AOS1869SUPP}.