Quantile Smoothing Splines
Roger W. Koenker, Pin Ng, Stephen L. Portnoy · Biometrika · 1994
Although nonparametric regression has traditionally focused on the estimation of conditional mean functions, nonparametric estimation of conditional quantile functions is often of substantial practical interest.We explore a class of quantile smoothing splines, which are defined as solutions to .1 min B>,(y/ -*(*/)) + XCj" \s"^)\ p ^) X ' p 2 e G J g 6 9 with p x (u) = (x-I(u < 0))w and p>\.For the particular choices p=\ and p=°° and g= [geC l [0,l]'-(J \g'\x)\ p dx) 1/p <°o}, we show that solutions, g, are parabolic splines, i.e. piecewise quadratics, on the mesh [x\, ..., x n ), and may be computed by standard / j -type linear programming techniques.At X = 0, g interpolates the X th quan- tiles at the distinct design points, and for X sufficiently large g is the linear regression quantile fit (Koenker and Bassett(1978)) to the observations.In fact, the entire path of solutions, in the penalty parameter X, may be efficiently computed by parametric linear programming methods.For a somewhat more general class of quantile smoothing splines we establish that (Ex(g(X)-go(X))-) =O p (n (logn) ), is achievable, under mild conditions on the true conditional quantile function of Y given X.Finally we note that the approach may be easily adapted to impose monotonicity and/or convexity constraints on the fitted function.
Cited by 635 works. doi:10.1093/biomet/81.4.673 Open-access PDF
Authors
References
- A Practical Guide to Splines
- Estimating the Dimension of a Model
- Generalized Additive Models.
- Spline Models for Observational Data.
- The New S Language.
- Smoothing noisy data with spline functions
- Smoothing reference centile curves: The lms method and penalized likelihood
- Applied Nonparametric Regression
- Spline Smoothing and Nonparametric Regression.
- Flexible Parsimonious Smoothing and Additive Modeling