fields (version 3.5)

Tps: Thin plate spline regression

Description

Fits a thin plate spline surface to irregularly spaced data. The smoothing parameter is chosen by generalized cross-validation. The assumed model is additive Y = f(X) +e where f(X) is a d dimensional surface. This is a special case of the spatial process estimate.

Usage

Tps(x, Y, m = NULL, p = NULL, scale.type = "range", ...)

Value

A list of class Krig. This includes the predicted surface of fitted.values and the residuals. The results of the grid search minimizing the generalized cross validation function is returned in gcv.grid. Please see the documentation on Krig for details of the returned arguments.

References

See "Nonparametric Regression and Generalized Linear Models" by Green and Silverman. See "Additive Models" by Hastie and Tibshirani.

Details

A thin plate spline is result of minimizing the residual sum of squares subject to a constraint that the function have a certain level of smoothness (or roughness penalty). Roughness is quantified by the integral of squared m-th order derivatives. For one dimension and m=2 the roughness penalty is the integrated square of the second derivative of the function. For two dimensions the roughness penalty is the integral of

(Dxx(f))**22 + 2(Dxy(f))**2 + (Dyy(f))**22

(where Duv denotes the second partial derivative with respect to u and v.) Besides controlling the order of the derivatives, the value of m also determines the base polynomial that is fit to the data. The degree of this polynomial is (m-1).

The smoothing parameter controls the amount that the data is smoothed. In the usual form this is denoted by lambda, the Lagrange multiplier of the minimization problem. Although this is an awkward scale, lambda =0 corresponds to no smoothness constraints and the data is interpolated. lambda=infinity corresponds to just fitting the polynomial base model by ordinary least squares.

This estimator is implemented by passing the right generalized covariance function based on radial basis functions to the more general function Krig. One advantage of this implementation is that once a Tps/Krig object is created the estimator can be found rapidly for other data and smoothing parameters provided the locations remain unchanged. This makes simulation within R efficient (see example below). Tps does not currenty support the knots argument where one can use a reduced set of basis functions. This is mainly to simplify and a good alternative using knots would be to use a valid covariance from the Matern family and a large range parameter.

See Also

Krig, summary.Krig, predict.Krig, predict.se.Krig, plot.Krig, surface.Krig, sreg

Examples

#2-d example 

fit<- Tps(ozone$x, ozone$y)  # fits a surface to ozone measurements. 

set.panel(2,2)
plot(fit) # four diagnostic plots of  fit and residuals. 
set.panel()

summary(fit)

# predict onto a grid that matches the ranges of the data.  

out.p<-predict.surface( fit)
image( out.p) 
surface(out.p) # perspective and contour plots of GCV spline fit 
# predict at different effective 
# number of parameters 
out.p<-predict.surface( fit,df=10)

#1-d example 
out<-Tps( rat.diet$t, rat.diet$trt) # lambda found by GCV 
plot( out$x, out$y)
lines( out$x, out$fitted.values)

# 
# compare to the ( much faster) one spline algorithm 
#  sreg(rat.diet$t, rat.diet$trt) 
# 

# Adding a covariate to the fixed part of model
# Note: this is a fairly big problem numerically (850+ locations)

Tps( RMprecip$x,RMprecip$y, Z=RMprecip$elev)-> out

surface( out, drop.Z=TRUE)
US( add=TRUE, col="grey")

#
# simulation reusing Tps/Krig object
#
fit<- Tps( rat.diet$t, rat.diet$trt)
true<- fit$fitted.values
N<-  length( fit$y)
temp<- matrix(  NA, ncol=50, nrow=N)
sigma<- fit$shat.GCV
for (  k in 1:50){
ysim<- true + sigma* rnorm(N) 
temp[,k]<- predict(fit, y= ysim)
}
matplot( fit$x, temp, type="l")


# 
#4-d example 
fit<- Tps(BD[,1:4],BD$lnya,scale.type="range") 

# plots fitted surface and contours 
# default is to hold 3rd and 4th fixed at median values 

surface(fit)