A study of nonparametric regression of error distribution in linear model based on \(L_ 1\)-norm (Q1894032)

We consider a linear model \[ y_i= {\mathbf x}_i' \beta+e_i, \quad i=1,2, \dots, \tag{1} \] where \({\mathbf x}_i'\)s are \(p\) \((\geq 1)\)-dimensional known vectors and \(\beta\) \((\in R^p)\) is an unknown regression coefficient vector. The errors \(e_i\) are assumed to be i.i.d. r.v.'s with a common unknown density function \(f(x)\), and \[ E(e_1)= 0,\quad 0<Var (e_1)= E(e^2_1) <\infty. \tag{2} \] It is frequently assumed that \(e_1\) has a normal distribution \(N(0, \sigma^2)\) in usual regression analysis. Then, an estimator of \(\beta\) based on \((x_1,y_1), \dots, (x_n,y_n)\) is obtained by the Least Squares method. Recently, the search work for robust procedures in statistical data analysis has generated considerable interest in developing statistical methods based on Least Absolute Deviations (LAD) estimators, which use the \(L_1\)-norm rather than the \(L_2\)-norm. The LAD estimator \(\widehat\beta\) of \(\beta\) in the model (1) is defined as a Borel measurable solution of the minimization problem: \[ \sum^n_{i=1} |y_i- {\mathbf x}_i' \widetilde \beta|=\min_{\beta \in R^p} \sum^n_{i=1} |y_i- {\mathbf x}_i' \beta|\tag{3} \] under the condition (4) \(med(e_1)=0\). We propose a nonparametric method for estimating an unknown error distribution function \(f(x)\) based on the LAD estimator in the general linear model (1) with condition (4), and prove that the nonparametric estimators have not only weak consistency, but also strong consistency. The asymptotic normality of the nonparametric estimator is also considered. The most difficulty in our study is: the estimator we propose here is based on residuals which are not independent, and also do not follow any fixed rules like those in dependent variables, say mixing variables.