English

Linear regression estimation in non-linear single index models

Statistics Theory 2017-12-19 v2 Statistics Theory

Abstract

In this article, we consider the problem of estimating the index parameter α0\alpha_0 in the single index model E[YX]=f0(α0TX)E[Y |X] = f_0(\alpha_0^T X) with f0f_0 the unknown ridge function defined on R\mathbb{R}, XX a d-dimensional covariate and YY the response. We show that when XX is Gaussian, then α0\alpha_0 can be consistently estimated by regressing the observed responses YiY_i, i=1,...,ni = 1, . . ., n on the covariates X1,...,XnX_1, . . ., X_n after centering and rescaling. The method works without any additional smoothness assumptions on f0f_0 and only requires that cov(f0(α0TX),α0TX)0cov(f_0(\alpha_0^T X),\alpha_0^TX) \neq 0, which is always satisfied by monotone and non-constant functions f0f_0. We show that our estimator is asymptotically normal and give the expression with its asymptotic variance. The approach is illustrated through a simulation study.

Keywords

Cite

@article{arxiv.1612.07704,
  title  = {Linear regression estimation in non-linear single index models},
  author = {Fadoua Balabdaoui and Gian-Andrea Thanei},
  journal= {arXiv preprint arXiv:1612.07704},
  year   = {2017}
}

Comments

The authors were made aware that a similar result already exists in the literature: "A generalized linear model with Gaussian regressor variables" (Brillinger, 1983)

R2 v1 2026-06-22T17:32:38.608Z