Envelope-based sparse partial least squares (Q2176612)

A multivariate linear regression model is considered \[Y=\mu+\beta^\top(X-\mu_X)+\epsilon,\] where \(Y\in\mathbb{R}^r\) is the response vector, and \(X\in\mathbb{R}^p\) is the stochastic predictor vector with mean \(\mu_X\) and finite 2nd moments. The errors \(\epsilon\in\mathbb{R}^r\) are independent of \(X\), and have zero mean and finite 2nd moments. The intercept and regression coefficients are unknown and denoted by \(\mu\in\mathbb{R}^r\) and \(\beta\in\mathbb{R}^{p\times r}.\) Sparse partial least squares (SPLS) estimators of \(\beta\) are widely used as a method that performs dimension reduction and variable selection, especially in high-dimensional setting, where \(p/n\) does not tend to zero as \(n\to\infty\). In the paper under review, a noval version of SPLS, called the envelope-based SPLS, is proposed using a connection between envelope models (see, e.g. [\textit{S. De Jong}, ``SIMPLS: an alternative approach to partial least squares regression'', Chemom. Intell. Lab. Syst. 18, 251--263 (1993; \url{doi:10.1016/0169-7439(93)85002-X})]) and partial least squares. The authors establish the consistency, asymptotic normality, and the selection consistency of the estimator. The oracle property is proven as well, which means that the estimator correctly selects the inactive predictors with probability tending to 1, and estimates the coefficients of the active predictors with the same efficiency as if the true model were known. The large-sample scenario and high-dimensional scenario are both considered. Also the paper develops the envelope-based SPLS estimators in the context of GLMs, and discusses the consistency, oracle property and asymptotic normality of the estimators. Simulations show that the envelope-based SPLS estimator has better variable selection and prediction performance over the SPLS estimator in [\textit{H. Chun} and \textit{S. Keleş}, J. R. Stat. Soc., Ser. B, Stat. Methodol. 72, No. 1, 3--25 (2010; Zbl 1411.62184)].