GRID: a variable selection and structure discovery method for high dimensional nonparametric regression (Q2196249)

In the paper, building on the work of \textit{J. Lafferty} and \textit{L. Wasserman} [Ann. Stat. 36, No. 1, 28--63 (2008; Zbl 1132.62026)], a new method, called the \textit{Gradient relevant identification of derivatives} (or the GRID) method, for simultaneous nonlinear variable selection and estimation of the low dimensional structure of the regression function under sparsity is developed. It combines empirical likelihood based marginal testing with the local linear estimation machinery to select the relevant variables. A simple graphical tool for identifying the low dimensional nonlinear structure of the regression function is provided. Theoretical results establish consistency of variable selection and structure discovery. The considered type of sparsity allows for interactions of arbitrarily high (but bounded) order that need not be specified a priori. Finite sample properties of the GRID are investigated in a moderately large simulation study. A supplementary material contains an algorithmic representation of the GRID method.