8
$\begingroup$

The following comes from a YouTube video: Robustness in Statistics, which I have tried to quote verbatim.

In Biology and Medicine these procedures are extremely popular, and I don't know why. They're useless, in my opinion. You can have, like, a non-parametric t-test, I think that's called a Wilcoxon. And then there's a Spearman rank correlation. And then, I have to look these up every time because I never use them, the Kruskall-Wallis and the Mann-Whitney U. So these are all "non-parametric" procedures that you use when you have screwy data. And when I was a bio-statistician I used these all the time, because I didn't know any better. But now I know better. So all of these are basically just transforming your data into ranks. So its just sorting the people in terms of highest score to lowest score and then analyzing the sorted data. I don't like that idea because you have all the disadvantages of transformations, so you lose the original scale of the variable. And you're not actually modelling the data; you're modelling the ranks of the data. You should model the data. If linear models don't fit, then use a different model folks. So yeah, these other methods are old, outdated, and the only people who use them are doctors and biologists. Yeah... I guess that ends my controversial opinion for today. Although amongst statisticians I don't think it is controversial. It's just... anyway, moving on. Long story short, don't use them.

These statistics have limitations, as all statistical methods do. They also have strengths, as some statistical methods do.

Since Dustin Fife has some training in Statistics, his commentary might suggest a common opinion in statistics among those with some statistical training. Purportedly he teaches courses in Statistics, and therefore has a tangible influence on reaching future statisticians along with practicing engineers, scientists, technologists, and mathematicians.

But I suspect he learned these opinions from somewhere, and I also suspect these views belong to a minority.

So to my question:

are there any surveys of the opinions of statisticians, preferably in the last 5 years, on the usefulness of rank-based classical nonparametric statistics?

Improve this question
$\endgroup$
7
  • 5
    $\begingroup$ What is the merit of a random person's opinion on the Internet? Is that person a world-leading statistician? A senior figure in some international statistical association? I strongly suspect that someone like Wasserman won't be throwing his All of Nonparametric Statistics book in the recycle bin just yet and ASA won't stop giving awards non-parametric Statistics in the immediate future... $\endgroup$
    – usεr11852
    Commented Dec 4, 2021 at 22:59
  • 3
    $\begingroup$ Even if Dustin Fife were the highest authority on the subject matter, outstripping Larry Wasserman in awards and accolades, I would still reject the cited paragraph as rhetorical nonsense. It exemplifies misinformation in someone with some statistical training, and who purportedly teaches courses in statistics. The real substance of my question is to find out how pervasive this misinformation is among statisticians. $\endgroup$
    – Galen
    Commented Dec 4, 2021 at 23:31
  • 3
    $\begingroup$ I wonder if you want to reword the question again to something like, "Are rank-based procedures useless?" I know stackexchange tends to be critical of opinion-based comments, but I think that would be a more useful discussion than simply asking if there are surveys out there. I wish we could have this discussion on my youtube channel, but YT seems to have blacklisted you from making comments. (Not my choice, btw. I didn't think your critical comments have been hostile and I would have loved to respond to them). I see your comments come to my inbox, but YT deletes them, for some reason. $\endgroup$
    – dfife
    Commented Dec 5, 2021 at 13:12
  • 2
    $\begingroup$ @user11852--to clarify, my criticism was specifically of classical rank-based nonparametric statistics (like Wilcoxen, Kruskall-Wallis, etc), not "modern nonparametrics" on which Wasserman's book is based. I routinely use/recommend bootstrapping and loess lines, but the classification I used in that video lumped those in a different category (as "modern robust methods"). $\endgroup$
    – dfife
    Commented Dec 5, 2021 at 19:48
  • $\begingroup$ @dfife I have accepted your edit as I think my language was harsher than intended. I think we still have disagreements about the usefulness of rank-based statistics, included Spearman's rho and others. I don't want to see it here because it would be off-topic from the question, but I would like to see a more substantive critique of these statistics from you in the future. $\endgroup$
    – Galen
    Commented Dec 5, 2021 at 20:08

0

Reset to default

Browse other questions tagged or ask your own question.