robLoc/robScale auxiliary/initial scale estimation

[filiporn]

Hi,

The paper states that "small-sample correction factors which make MADn ...
unbiased" are used. Dr Rousseeuw provided the original source code which confirms that the factors from this paper were used.

An implementation of this could look like

madfc <- function(x) {
x <- x[!is.na(x)] # Remove NAs
n <- length(x)

bn <- if (n == 2) {
1.196
} else if (n == 3) {
1.495
} else if (n == 4) {
1.363
} else if (n == 5) {
1.206
} else if (n == 6) {
1.200
} else if (n == 7) {
1.140
} else if (n == 8) {
1.129
} else if (n == 9) {
1.107
} else if (n > 9) {
n / (n - 0.8)
} else {
1
}
mad_fc <- mad(x) * bn
return(mad_fc)
}

And in robLoc/robScale "mad" would be replaced by "madfc".

Best, Filip

[aadler]

Thank you, @filiporn. Sorry about the LONG delay in responding. I will review the paper and (probably) update the function. As it is, one of these days, I want to do some testing as to how many observations are needed to revert to using standard estimators. The authors mention about 8, if I recall correctly.

[aadler]

@filiporn, you are correct. R&V specifically use the MAD_n version in their 2002 paper, which is the one with the constants found in C&R 1992. So please check the devel branch where I have implemented MAD_n as mdm (median deviation from the median per R&V 2002) and now use it in robLoc and robScale. I will need to wait until CRAN returns from holiday break to push it, though. It will be part of the 3.0.0 release. Thank you! PS, I've thanked your GitHub handle in the NEWS. Let me know (before 2026-01-07 if possible :) ) if you would like to be referred to by a different name. Thanks!!!

[aadler]

I renamed the new function madn to more closely follow both papers.

[aadler]

I've made some pretty significant changes in the package construction. Also, I think I may want to hold off pushing the new version until I write/submit the paper describing the larger Monte-Carlo simulation.

[filiporn]

Hi! I am glad you found my comment useful. :)