Kernel Spline Regression
by
John Braun
University of Western Ontario

In this talk, a method of combining spline regression and kernel smoothing is proposed. A simulation study shows that the resulting curve estimators can succeed in capturing the main features of the true curve more effectively than local polynomial regression when the curvature rapidly changes. Kernel splines also exhibit somewhat less sensitivity to choice of bandwidth than local polynomial regression. In many situations, the knots can be chosen manually with some ease. This is an advantage over variable bandwidth and variable order approaches in kernel polynomial regression which cannot be handled quite as easily. When automatic knot selection is required, the knot deletion method seems to work well. A direct-plug-in bandwidth selector is available as well. Some illustrative examples will also be presented.

Date received: September 19, 2003