Some Problems of Statistical Estimation of the Regression Curve

From MaRDI portal





Two different formulations are known for the problem of the estimation of the regression curve. In the first of these, it is required to estimate the function \(\lambda(x)= {\mathbf E} (Y\mid X=x)\), using independent observations of the pairs \((X_ 1, Y_ 1),\dots, (X_ N,Y_ N)\), having the same distribution function \(F\), where the pair \((X,Y)\) also has the distribution function \(F\). In the second formulation the choice of the points \(X_ 1,\dots,X_ N\) of observations \(Y_ 1(X_ 1),\dots, Y_ N(X_ N)\) is determined by the observer. The present paper deals with some questions concerning the problem considered in the first formulation. Assuming that \((X,Y)\) is a two-dimensional random variable with probability density \(f(x,y)\), we construct and study statistical estimators of the regression curve \[ \begin{aligned} \lambda(x) &={\mathbf E} (Y\mid X=x)= \gamma(x)/ p(x)\\ \text{and its derivative} \lambda'(x) &= [p(x) \gamma'(x)- \gamma(x) p'(x)]/ p^ 2(x), \end{aligned} \] where \(p(x)= \int f(x,y)dy\) and \(\gamma(x)= \int yf(x,y)dy\). The estimators introduced by us for the regression curve \(\lambda(x)\) and its derivative are constructed taking into account certain a priori assumptions on the smoothness of the functions \(p(x)\) and \(\gamma(x)\) and are distinguished by a high rate of uniform convergence to the true values of the estimated functions (within the limits of some bounded segment \(U\)). Here it seems reasonable to note that the smoothness of the functions \(p(x)\) and \(\gamma(x)\) may be quite different. We shall not dwell on the proof of this fact, altough it is not complicated. Further, we study estimators of zeros and extrema of the function \(\lambda(x)\).











This page was built for publication: Some Problems of Statistical Estimation of the Regression Curve

Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q3974386)