[R] package quantreg behaviour in weights in function rq,

[Václav Varvařovský]

Dear all,

I am having v.4.36 of Quantreg package and I noticed strange behaviour when
weights were added. Could anyone please explain me what if the results are
really strange or the behavioiur is normal. As an example I am using dataset
Engel from the package and my own weights.

x<-engel[1:50,1]
y<-engel[1:50,2]
w<-c(0.00123, 0.00050, 0.00126, 0.00183, 0.00036, 0.00100,
0.00122, 0.00133, 0.01208, 0.00126, 0.00102, 0.00183,
0.00063, 0.00134, 0.00084, 0.00087, 0.00118, 0.00894,
0.00105, 0.00154, 0.02829, 0.00095, 0.05943, 0.07003,
0.00692, 0.03610, 0.00316, 0.06862, 0.00439, 0.08974,
0.01960, 0.00185, 0.00348, 0.03597, 0.00210, 0.03929,
0.03535, 0.01463, 0.02254, 0.00089, 0.01495, 0.00178,
0.00351, 0.10338, 0.13662, 0.00157, 0.07689, 0.07304,
0.00194, 0.00142)

windows(width = 8, height = 7)
u1<-rq(y~x,tau=c(1:10/10-0.05),weights=w)
#note that by taking weights = 500*w the points
#are all moving down (i thought it should have been invariant to the
magnitude of weights) when all the weights remaint the same
plot(x,y,ylim=c(-1000,max(y)))
points(x,u1$fitted.values[,3],col="blue",cex=0.5)
#weighted - fitted values nearly match original values

windows(width = 8, height = 7)
u1<-rq(y~x,tau=c(1:10/10-0.05))
plot(x,y,ylim=c(-1000,max(y)))
points(x,u1$fitted.values[,3],col="blue",cex=0.5)
#unweighted - fitted values form a line

Why the weighted quantile regression does not produce a line?

Thank you for answers.

[[alternative HTML version deleted]]

______________________________________________
R-h...@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-help
PLEASE do read the posting guide http://www.R-project.org/posting-guide.html
and provide commented, minimal, self-contained, reproducible code.

[roger koenker]

Václav,

I think that the technical term for this is snafu. You are right of
course
that the magnitude of the weights should have no impact on the fitted
values. The good news is that the coefficient estimates satisfy this
obvious principle, but unfortunately the fitted.values component is
being
computed with the wrong form of the X matrix, so your plots are
misaligned. The immediate work around is to compute fitted values
from the original data and the coef component of the fit rather than
relying on the fitted.values component. Thanks for pointing this out,
I'll try to submit a repaired version of the package later in the day.

Roger

url: www.econ.uiuc.edu/~roger Roger Koenker
email rkoe...@uiuc.edu Department of Economics
vox: 217-333-4558 University of Illinois
fax: 217-244-6678 Urbana, IL 61801