[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: RE: winzorized mean comparison
. findit winsor
suggests not. But then -findit winsor- doesn't show up
-winsor- on SSC: presumably SSC was inaccessible the last
time the search engine did its search.
You don't say so, but I guess your, or your colleague's,
concern is getting P-values for some tests.
Without knowing what it is
I guess there is a small literature somewhere giving
rules of thumb and/or approximate P-values for some
situations. Rand Wilcox of UCLA or USC is one name who springs
to mind. But who knows if the situations covered include
Winsoring is one thing that some people do when they are
worried about tails. I would suggest some simple tactics:
1. Push your Winsorised variables through whatever
standard commands you want to run. Also push the
un-Winsorised variables through same.
2. If the results are scientifically similar, all
you need do is report the analysis on the unWinsorised
together with a comment that the tails are not
wagging the analysis, as shown by similar results
with Winsorising. I don't know if some people would
argue that you should lose a few d.f. for making this
3. If the results are scientifically very different,
worrying starts. If the Winsorising arises because
of _one_ wobbly tail, perhaps you should be transforming
or using e.g. a generalised linear model with non-identity
link. If the Winsorising arises because of approximately
symmetric but also long-tailed distributions, there
are also transformations for this case. In practice that
seems less common.
Incidentally, I guess the spelling could be "Winsorize"
or "Winsorize" depending on which syntax scholar you
follow, but "Winzor" for "Winsor" is just a typo. The
procedure is named for the engineer-turned-biostatistician
Charles P. Winsor (1895--1951).
Chapman, Ben P
> A colleague in biostatistics has a Stata inquiry that's
> stumped me. I know there's a program for winsorizing means,
> but he is curious to know if stata has any procedure for
> comparing two or more independent groups with trimmed or
> winsorized means, rather than conventional means as in a
> t-test. Apropros of the recent discussion about FAQs, I have
> indeed searched and see nothing on this. Any thoughts much
* For searches and help try: