Nonparametric Conditional Estimation

: Many nonparametric regression techniques (such as kernels, nearest neighbors, and smoothing splines) estimate the conditional mean of Y given X = x by a weighted sum of observed Y values, where observations with X values near x tend to have larger weights. In this report the weights are taken to represent a finite signed measure on the space of Y values. This measure is studied as an estimate of the conditional distribution of Y given X= x. From estimates of the conditional distribution, estimates of conditional means, standard deviations, quantiles and other statistical functionals may be computed. Chapter 1 illustrates the computation of conditional quantiles and conditional survival probabilities on the Stanford Heart Transplant data. Chapter 2 contains a survey of nonparametric regression methods and introduces statistical metrics and von Mises' method for later use. Chapter 3 proves some consistency results. The estimated conditional distribution of Y is shown to be consistent in the following sense: the Prohorov distance between the estimated and true conditional distributions converges in probability to zero. The required conditions are: that the distribution of Y given X = x vary continuously with x, that the weights regarded as a measure on the X space converge in probability to a point mass at x, and that a measure of the effective local sample size tend to infinity in probability. A slight strengthening of the conditions allows one to establish almost sure consistency. Consistency of Prohorov-continuous (i.e. robust) functionals follows immediately. In the above, the X and Y spaces are complete separable metric spaces. In case Y is the real line, weak and strong consistency results are established for the Kolmogorov-Smirnov and the Vasserstein metrics under stronger conditions. Chapter 4 provides conditions under which the suitably normalized errors in estimating the conditional distribution of Y have a Brownian limit.