fitdiag: Robust Fit Diagnostics

In kloke/hbrfit: High breakdown rank-based (HBR) estimation for linear models

Robust Fit Diagnostics

Description

Diagnostics to compare two fits of a linear model. The comparison (tdbeta) is a measure of total difference between the fits while the comparison (cfit) is measure of the difference between the fitted (predicted) values. Comparisons between any two of the following fits can be made: Wilcoxon (WIL), least squares (LS), HBR (HBR), and least trim squares.

Usage

fitdiag(x, y, est = c("WIL", "HBR"), ...)

Arguments

x	design matrix.
y	response vector.
est	character vector of length 2. estimates to use. default is est = c("WIL", "HBR"))
...	additional arguments passed to rfit.

Details

The diagnostic tdbeta uses the benchmark bmtd. Values exceeding bmtd indicate the fits differ. The diagnostic cfit uses the benchmark bmcf; for each case_i, cfit_i exceeding bmcf in absolute value indicates the ith fitted values differ. These are only benchmarks, not critical values. A useful plot is cfit versus case. The user should look for large gaps in this plot.

Value

tdbeta	Total difference.
cfit	Casewise differences in fits.
bmtd	Benchmark for tdbeta.
bmcf	Benchmark for cfit.
est	The two fits being compared.

Author(s)

Jeff Terpstra, Joe McKean, John Kloke

References

Hettmansperger, T.P. and McKean J.W. (2011), Robust Nonparametric Statistical Methods, 2nd ed., New York: Chapman-Hall.

Examples

# Case of no outliers
set.seed(35443092)
x<-rnorm(20);y<-rnorm(20)
dwl <- fitdiag(x,y,est=c("WIL","LS"))
tdbetas <- round(c(dwl$tdbeta,dwl$bmtd),digits=2)
tdbetas
Case<-1:length(y)
plot(dwl$cfit~Case,ylab='Difference in fits',ylim=c(-dwl$bmcf,dwl$bmcf))
title('Simulated Normal Data')
abline(h=c(-dwl$bmcf,dwl$bmcf))  # benchmark bars
# Case where there is a gap
dwilhbr = fitdiag(stars$temperature,stars$light,est=c("WIL","HBR"))
tdbetas <- round(c(dwilhbr$tdbeta,dwilhbr$bmtd),digits=2)
tdbetas
Case <- 1:length(stars$temperature)
plot(dwilhbr$cfit~Case,ylab='Difference in fits')
title('Stars Data with Six Outliers')

kloke/hbrfit documentation built on Nov. 17, 2023, 2:33 p.m.