Significance testing in non-sparse high-dimensional linear models (Q1616315)

Following up on prior work by the auhtors [J. Am. Stat. Assoc. 113, No. 524, 1583--1600 (2018; Zbl 1409.62139)], the authors are concerned with statistical methods to evaluate the importance of variables in linear models. In particular, they treat high-dimensional settings where the number $p$ of regressors may exceed the sample size $n$. The so-called CorrT test is introduced for testing the (global) null hypothesis that the vector $\beta \in \mathbb{R}^p$ of regression coefficients of interest equals some given vector $\beta_{0} \in \mathbb{R}^p$. Its test statistic $T_n$ is based on self-adaptive and sparsity-adaptive estimators of nuisance parameters of the model. The authors provide methods for an efficient implementation of the proposed test, and they study its theoretical properties (asymptotic size and power as $n$ and $p$ tend to infinity). Finally, the usage of the proposed methods is illustrated on simulated and real data.