Confidence Intervals Based on Local Linear Smoother (Q4416163)

This article deals with confidence intervals for an unknown regression function based on the asymptotic normal distribution of the local linear smoother. Let \(\{(X_{i},Y_{i})\}^{n}_{i=1}\) be \(n\) independent pairs of observations where \(Y_{i}\) are the responses of design points \(X_{i}\), for \(i=1,\dots,n\), according to the regression model \(Y_{i}=m(X_{i})+\epsilon_{i}\), where \(m\) is an unknown regression function with \(\mathcal L\) being its support, \(\epsilon_{i}\) are independent errors with zero mean and variance function \(V(X_{i})\), and the design points \(X_{i}\in R\) have density \(f\). Let \(K\) be a kernel function which is a probability density itself and \(h\) be the smoothing bandwidth. Let us define \(V(x)=E[(Y-m(x))^2\mid X=x]\). The local linear regression estimator for \(m(x)\) at any given \(x\in{\mathcal L}\) is \(\hat m(x)=\sum_{i=1}^{n}W_{i} Y_{i}/\sum_{i=1}^{n}W_{i}\), where \[ W_{i}=K_{h}(x-X_{i})\{s_{n,2}- (x-X_{i})s_{n,1}\},\quad s_{n,l}=\sum_{i=1}^{n}K_{h}(x-X_{i})(x- X_{i})^{l}\text{ for }l=0,1,2, \] and \(K_{h}(\cdot)=h^{-1}K(\cdot/h)\). Let \(\hat V(x)\) and \(\hat f(x)\) be some consistent estimators of \(V(x)\) and \(f(x)\), and let \(z_{\alpha}\) be the \((1+\alpha)/2\)-quantile of \(N(0,1)\). Then a normal approximation based confidence interval for \(m(x)\) with nominal coverage \(\alpha\) is \[ I_{\alpha}=\hat m(x)\pm z_{\alpha} \sqrt{v(x/h)\hat V(x)/[nh\hat f(x)]}, \] where \[ v(x/h)= \left(\int_{-1}^{x/h}[\alpha_2(x/h)-u\alpha_1(x/h)]^2K^2(u)du\right) [\alpha_0(x/h)\alpha_2(x/h)-\alpha_1^2(x/h)]^{-2}, \] \(\alpha_{l}(x/h)=\int_{-1}^{x/h}u^{l}K(u) du\). The coverage accuracy of the proposed confidence intervals in the interior and in the boundary are evaluated separately by using Edgeworth expansions. The present study shows that the coverage error near the boundary is of a larger order of magnitude than that in the interior of the support.