Universal weighted kernel-type estimators for some class of regression models (Q2227200)

The authors study nonparametric regression problems of the form \[ Y_i = f(X_i) + \xi_i, \quad i=1, \dots, n, \] for univariate responses \(Y_1, \ldots, Y_n\) and \([0, 1]\)-valued, random design points \(X_1, \dots, X_n\), which form a triangular array of random variables and are not necessarily stochastically independent or identically distributed. The error terms \((\xi_i)_{1 \leq i \leq n}\) are assumed to be (conditionally) centered and uncorrelated, but not necessarily stochastically independent or identically distributed. The target of statistical inference is the unknown random regression function \(f\), which can depend on \(\omega \in \Omega\), such that \(f(t) \equiv f(\omega, t)\) for \(t \in [0, 1]\). A class of kernel-based weighted least squares estimators for \(f\) is proposed and studied. In particular, under certain conditions, explicit upper bounds for their rate of uniform convergence in probability are derived as the sample size \(n\) tends to infinity. This result is applied to the problem of estimating the mean of \(f\). Furthermore, pointwise asymptotic normality of the proposed estimators is proven, with applications to the construction of (asymptotic) pointwise confidence intervals. In the special case of a nonrandom \(f\), stochastically independent and identically distributed (i.i.d.) design points, and i.i.d. centered random errors, the proposed estimators are compared with Nadaraya-Watson estimators in terms of (pointwise) biases and variances. Finally, two concrete examples are presented.