fields (version 16.3)

QTps: Robust and Quantile smoothing using a thin-plate spline

Description

This function uses the standard thin plate spline function Tps and a algorithm based on psuedo data to compute robust smoothers based on the Huber weight function. By modifying the symmetry of the Huber function and changing the scale one can also approximate a quantile smoother. This function is experimental in that is not clear how efficient the psuedo-data algorithm is acheiving convergence to a solution.

Usage

QTps(x, Y, ..., f.start = NULL, psi.scale = NULL, C = 1, alpha = 0.5, Niterations = 100,
               tolerance = 0.001, verbose = FALSE)
QSreg(x, Y, lambda = NA, f.start = NULL, psi.scale = NULL, 
    C = 1, alpha = 0.5, Niterations = 100, tolerance = 0.001, 
    verbose = FALSE)

Value

A Krig object with additional components:

yraw

Original Y values

conv.info

A vector giving the convergence criterion at each iteration.

conv.flag

If TRUE then convergence criterion was less than the tolerance value.

psi.scale

Scaling factor used for the psi.wght function.

value

Value of alpha.

Arguments

x

Locations of observations.

Y

Observations

lambda

Value of the smoothing parameter. If NA found by an approximate corss-validation criterion.

...

Any other arguments to pass to the Tps function, which are then passed to the Krig function. give.warnings =FALSE can be used to turn off pesky warnings when not important (see example below).

C

Scaling for huber robust weighting function. (See below.) Usually it is better to leave this at 1 and just modify the scale psi.scale according to the size of the residuals.

f.start

The initial value for the estimated function. If NULL then the constant function at the median of Y will be used. NOTE: This may not be a very good starting vector and a more robust method would be to use a local robust smoother.

psi.scale

The scale value for the Huber function. When C=1, this is the point where the Huber weight function will change from quadratic to linear. Default is to use the scale .05*mad(Y) and C=1 . Very small scales relative to the size of the residuals will cause the estimate to approximate a quantile spline. Very large scales will yield the ordinary least squares spline.

alpha

The quantile that is estimated by the spline. Default is .5 giving a median. Equivalently this parameter controls the slope of the linear wings in the Huber function 2*alpha for the positive wing and 2*(1-alpha) for the negative wing.

Niterations

Maximum number of interations of the psuedo data algorithm

tolerance

Convergence criterion based on the relative change in the predicted values of the function estimate. Specifically if the criterion mean(abs(f.hat.new - f.hat))/mean(abs(f.hat)) is less than tolerance the iterations re stopped.

verbose

If TRUE intermediate results are printed out.

Author

Doug Nychka

Details

These are experimental functions that use the psuedo-value algorithm to compute a class of robust and quantile problems. QTps use the Tps function as its least squares base smoother while QSreg uses the efficient sreg for 1-D cubic smoothing spline models. Currently for the 1-d spline problem we recommend using the (or at least comparing to ) the old qsreg function. QSreg was created to produce a more readable version of the 1-d method that follows the thin plate spline format.

The Thin Plate Spline/ Kriging model through fields is: Y.k= f(x.k) = P(x.k) + Z(x.k) + e.k

with the goal of estimating the smooth function: f(x)= P(x) + Z(x)

The extension in this function is that e.k can be heavy tailed or have outliers and one would still like a robust estimate of f(x). In the quantile approximation (very small scale parameter) f(x) is an estimate of the alpha quantile of the conditional distribution of Y given x.

The algorithm is iterative and involves at each step tapering the residuals in a nonlinear way. Let psi.wght be this tapering function then given an initial estimate of f, f.hat the new data for smoothing is

Y.psuedo<- f.hat + psi.scale* psi.wght( Y - f.hat, psi.scale=psi.scale, alpha=alpha) A thin plate spline is now estimated for these data and a new prediction for f is found. This new vector is used to define new psuedo values. Convergence is achieved when the the subsequent estimates of f.hat do not change between interations. The advantage of this algorithm is at every step a standard "least squares" thin plate spline is fit to the psuedo data. Because only the observation vector is changing at each iteration Some matrix decompositions need only be found once and the computations at each subsequent iteration are efficient. At convergence there is some asymptotic theory to suggest that the psuedo data can be fit using the least squares spline and the standard smoothing techinques are valid. For example one can consider looking at the cross-validation function for the psuedo-data as a robust version to select a smoothing parameter. This approach is different from the weighted least squared algorithm used in the qsreg function. Also qsreg is only designed to work with 1-d cubic smoothing splines.

The "sigma" function indicating the departure from a pure quadratic loss function has the definition

qsreg.sigma<-function(r, alpha = 0.5, C = 1) 
     temp<- ifelse( r< 0, ((1 - alpha) * r^2)/C ,  (alpha * r^2)/C)
     temp<- ifelse( r >C,   2 * alpha * r - alpha * C, temp)
     temp<- ifelse( r < -C, -2 * (1 - alpha) * r - (1 - alpha) * C, temp)
     temp

The derivative of this function "psi" is

 qsreg.psi<- function(r, alpha = 0.5, C = 1)
     temp <- ifelse( r < 0, 2*(1-alpha)* r/C, 2*alpha * r/C )               
     temp <- ifelse( temp > 2*alpha, 2*alpha, temp)
     temp <- ifelse( temp < -2*(1-alpha), -2*(1-alpha), temp)
     temp

Note that if C is very small and if alpha = .5 then psi will essentially be 1 for r > 0 and -1 for r < 0. The key feature here is that outside a ceratin range the residual is truncated to a constant value. This is similar to the Windsorizing operation in classical robust statistics.

Another advantage of the psuedo data algotrithm is that at convergence one can just apply all the usual generic functions from Tps to the psuedo data fit. For example, predict, surface, print, etc. Some additional components are added to the Krig/Tps object, however, for information about the iterations and original data. Note that currently these are not reported in the summaries and printing of the output object.

References

Oh, Hee-Seok, Thomas CM Lee, and Douglas W. Nychka. "Fast nonparametric quantile regression with arbitrary smoothing methods." Journal of Computational and Graphical Statistics 20.2 (2011): 510-526.

See Also

qsreg

Examples

data(ozone2)
x<- ozone2$lon.lat
y<- ozone2$y[16,]



# Smoothing fixed at 50 df 
    look1<- QTps( x,y, psi.scale= 15, df= 50)

if (FALSE) {
# Least squares spline (because scale is so large)
    look2<- QTps( x,y, psi.scale= 100, df= 50)
#
    y.outlier<- y
# add in a huge outlier.
    y.outlier[58]<- 1e5
    look.outlier1<- QTps( x,y.outlier, psi.scale= 15, df= 50,
                           give.warnings= FALSE)
# least squares spline.
    look.outlier2<- QTps( x,y.outlier, psi.scale=100 , df= 50,
                           give.warnings= FALSE)
#
    set.panel(2,2)
    surface( look1)
    title("robust spline")
    surface( look2)
    title("least squares spline")
    surface( look.outlier1,  zlim=c(0,250))
    title("robust spline w/outlier") 
    points( rbind(x[58,]), pch="+")
    surface( look.outlier2, zlim=c(0,250))
    title("least squares spline w/outlier")
    points( rbind(x[58,]), pch="+")
    set.panel()
}
# some quantiles
look50 <- QTps( x,y, psi.scale=.5,)
look75 <- QTps( x,y,f.start= look50$fitted.values, alpha=.75)


# a simulated example that finds some different quantiles. 
if (FALSE) {
set.seed(123)
N<- 400
x<- matrix(runif( N), ncol=1)
true.g<- x *(1-x)*2
true.g<- true.g/ mean( abs( true.g))
y<-  true.g + .2*rnorm( N )

look0 <- QTps( x,y, psi.scale=10, df= 15)
look50 <- QTps( x,y, df=15)
look75 <- QTps( x,y,f.start= look50$fitted.values, df=15, alpha=.75)
}

if (FALSE) {
# this example tests the quantile estimate by Monte Carlo
# by creating many replicate points to increase the sample size. 
# Replicate points are used because the computations for the 
# spline are dominated by the number of unique locations not the 
# total number of points. 
set.seed(123)
N<- 80
M<- 200
x<- matrix( sort(runif( N)), ncol=1)
x<- matrix( rep( x[,1],M), ncol=1)

true.g<- x *(1-x)*2
true.g<- true.g/ mean( abs( true.g))
errors<- .2*(rexp( N*M) -1)
y<- c(matrix(true.g, ncol=M, nrow=N) + .2 *  matrix( errors, ncol=M, nrow=N))

look0 <- QTps( x,y, psi.scale=10, df= 15)
look50 <- QTps( x,y, df=15)
look75 <- QTps( x,y, df=15, alpha=.75)


bplot.xy(x,y, N=25)
xg<- seq(0,1,,200)
lines( xg, predict( look0, x=xg), col="red")
lines( xg, predict( look50, x=xg), col="blue")
lines( xg, predict( look75, x=xg), col="green")
}
if (FALSE) {
# A comparison with qsreg
  qsreg.fit50<- qsreg(rat.diet$t,rat.diet$con, sc=.5)
  lam<- qsreg.fit50$cv.grid[,1]
  df<- qsreg.fit50$cv.grid[,2]
  M<- length(lam)
  CV<-rep( NA, M)
  M<- length( df)
  fhat.old<- NULL
  for ( k in M:1){
     temp.obj<- QTps(rat.diet$t,rat.diet$con, f.start=fhat.old,  psi.scale=.5, tolerance=1e-6,
     verbose=FALSE, df= df[k],
     give.warnings=FALSE) 
     # avoids warnings from Krig search on lambda
     cat(k, " ")
     CV[k] <- temp.obj$Qinfo$CV.psuedo
     fhat.old<- temp.obj$fitted.values
  }
  plot( df, CV, type="l", lwd=2)
# psuedo data estimate
  points( qsreg.fit50$cv.grid[,c(5,6)], col="blue")
# alternative CV estimate via reweighted LS
  points( qsreg.fit50$cv.grid[,c(2,3)], col="red")
}