We report some topics on spline smoothing, that is, a method of nonparametric regression. First we describe one-dimensional spline smoothing. An estimator of the regression function is derived, and the methods of computing estimates and choosing the smoothing parameter are summarized. Next, in the case of several explanatory variables, we discuss a semiparametric regression model. In this model, a nonparametric function on one (or more) variable (s), and a linear function on the other variables, are fitted, and we assume the additivity on them. The estimators of both components are derived, and additionally this model is applied to some data of agricultural experiment. In the latter part we descrbe some asymptotic properties on semiparametric regression estimators. Rice (1986) pointed out that estimates of the regression coefficients in the parametric part were biased. In order to reduce that bias, two estimators were proposed. So we try to compare numerically by simulation to what extent the bias of these estimators is reduced. As a result, we find that the partial regression estimator has the effect of reducing the bias to some extent even if sample size is small, whereas the two-stage spline smoothing estimator is effective only if sample size is small.
View full abstract