Skip to content

[RsR] extracting weights from nlrob

2 messages · Andrew Halford, Andreas Ruckstuhl

Original
Andrew Halford
# 
Hi Listers,

I have been running a robust 2nd order polynomial regression and I want to
extract the weights information to see what datapoints are behaving as
outliers. Although I can see the summary data for the weights as part of
the output, I have not been able to find any help for extracting all the
individual weights. I would also like to be able to colour these points on
a scatterplot if possible.

I also cannot find any R packages that provide a way to calculate and
remove outliers in nonlinear regressions - along the lines of the paper by
Motulsky and Brown 2006.

Any advice on this would be most welcome.

Thanks for help.

Andy
Andreas Ruckstuhl
# 
Dear Andy

I would recommend to use lmrob for running polynomial regression. It can 
handle outliers on leverage points much better than nlrob.

library(robustbase)

set.seed(54362)
DAT <- data.frame(x=c(seq(-5,5,by=1), 8))
DAT$xq <- DAT$x^2
DAT$y <- 2 + 3*DAT$x + 1.2*DAT$xq + rnorm(nrow(DAT), sd=1.5)
DAT$y[nrow(DAT)] <- 60

DAT.lm <-  lm(y ~ x + xq, data=DAT)
summary(DAT.lm)
## Coefficients:
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)   5.9686     2.4849   2.402 0.039770 *
## x             2.1114     0.5750   3.672 0.005141 **
## xq            0.7117     0.1257   5.663 0.000308 ***
## ---
## Residual standard error: 6.41 on 9 degrees of freedom


DAT.rlm <-  lmrob(y ~ x + xq, data=DAT)
summary(DAT.rlm)
##             Estimate Std. Error t value Pr(>|t|)
## (Intercept)  1.48591    0.47660   3.118   0.0124 *
## x            2.72508    0.08822  30.890 1.91e-10 ***
## xq           1.24015    0.02458  50.457 2.37e-12 ***
## ---
## Robust residual standard error: 1.451

str(DAT.rlm)

(h.rw <- DAT.rlm$weights)
h.sel <- h.rw < 0.1

plot(y~x, data=DAT)
lines(DAT$x, predict(DAT.rlm), col="blue", lwd=2)
lines(DAT$x, predict(DAT.lm), col="red")
points(DAT$x[h.sel], DAT$y[h.sel], pch=16, col="orange")

## I don't know the paper by Motulsky and Brown 2006. I do not recommend
## removing outliers but just recommend identifying outliers and if
## possible correct typos and such things.



DAT.nlrob1 <-  nlrob(y ~ a + b*x + c*x^2, start=list(a=6,b=2,c=0.7),
                      data=DAT)
summary(DAT.nlrob1)
## Formula: y ~ a + b * x + c * x^2
## Parameters:
##   Estimate Std. Error t value Pr(>|t|)
## a   5.1419     2.0241   2.540  0.03169 *
## b   1.8283     0.4775   3.829  0.00403 **
## c   0.7177     0.1004   7.148 5.38e-05 ***
## ---
## Robust residual standard error: 4.521

## nlrob cannot cope wit outliers on leverage points!

## Hence
h.d <- (DAT$x-median( DAT$x))/mad(DAT$x)
DAT$wx <- 1/sqrt(1+8*pmax(0, h.d^2-1)/sqrt(2))
DAT.nlrob2 <-  nlrob(y ~ a + b*x + c*x^2, start=list(a=6,b=2,c=0.7) 
,data=DAT,
                      weights=DAT$wx)
summary(DAT.nlrob2)
## Parameters:
##   Estimate Std. Error t value Pr(>|t|)
## a  1.81836    0.74151   2.452   0.0366 *
## b  2.69350    0.16763  16.068 6.20e-08 ***
## c  1.19773    0.05919  20.236 8.19e-09 ***
## ---
## Robust residual standard error: 1.417

str(DAT.nlrob2)
DAT.nlrob2$w.r
  ## [1] 1.00000000 1.00000000 1.00000000 1.00000000 1.00000000 1.00000000
  ## [7] 1.00000000 0.88186039 1.00000000 1.00000000 1.00000000 0.04762281

## These are more reliable results.
## But be careful. This approach with weighting does not yield a
## high-breakdown estimator.


And by the way:
I guess you will be very careful with expressions like 'DAT$x^2' because 
of numerical issues, won't you.



All the best,
Andreas




Am 20.11.2012 09:05, schrieb Andrew Halford: