Optimal estimation of slope vector in high-dimensional linear transformation models (Q1755121): Difference between revisions

A sparse transformational regression model is considered where a response variable \(Y\) is related to predictor random variables \(X=(X_0,\ldots,X_p)^\top\) through the equation \(h(Y)=X^\top \beta+\epsilon\). Here \(h\) is an unknown strictly increasing function, \(\beta\) is a sparse slope vector of interest, and \(\epsilon\) is a noise term independent of \(X\) with unknown distribution. The Constrained Elastic Net (CENet) is proposed, a new method for performing simultaneous estimation and variable selection. CENet is the solution to a convex optimization problem and hence can be computed efficiently. Under a jointly elliptical distributional assumption on \((X, \epsilon)\), the CENet estimator achieves the same optimal rate of convergence as the best regression method in the high-dimensional sparse linear regression model. CENet does not require the specification of the function \(h\) and the noise distribution, and so is more robust compared to existing estimation and variable selection methods in the large-\(p\)-small-\(n\) setting, where \(n\) is the sample size. The robustness of CENet to both nonlinearity in the response and heavy-tailedness in the noise distribution is confirmed by the theoretical results and simulation studies. It is shown that there is a connection between CENet and existing nonlinear regression/multivariate methods, when the elliptical assumption on \((X, \epsilon)\) holds; in that case, an advantage of CENet is that it bypasses the need to explicitly estimate \(h\) when the primary interest is the estimation of \(\beta\).