You are doubly wrong Jeff. First of all, what you call a nonparametric procedure is a rank or permutation test based on ranks. If you were planning to look up the p-values in a table of the F-distribution, you will get wrong answers. If you were planning to use the permutation distribution of the ranks, then you must limit it to those permutations which are synchronous (this term and the concept are due to Pesarin whom you've chosen to debate). And why take ranks, when you can just as easily obtain a permutation test based on the original observations?
For an explantion of the methodology, please refer to http://users.oco.net/drphilgood/exchange.htm
I'd also recommend Pesarin's text on the subject.
Jeff Rasmussen <[log in to unmask]> wrote: No, the question asked was how to run a non-parametric test & the answer of
converting to ranks is a correct procedure (amongst various other
non-parametric procedures). You're presenting a different issue entirely.
Your statements below are correct that converting to ranks changes the
null-hypothesis, and therefore the ranks may show an interaction where the
raw data do not (or vice versa); and some other procedure may be preferable
to a rank-transformation.
----- Original Message -----
From: "Fortunato Pesarin"
To: "Jeff Rasmussen"
Sent: Friday, April 12, 2002 3:54 AM
Subject: Re: FW: non parametric problems
> Dear colleagues,
>
> the idea to convert data into ranks and then run a regular 3-way anova is
> incorrect. In fact the rank transformation may induce apparent
interactions
> where there are none, or may hidden interactions or even main effects
where
> they are present.
> One possibility is to run "syncronized permutations" as they are described
> in chapter 8 of my book: F. Pesarin, Multivariate Permutation tests with
> Applications in Biostatistics, Wiley, Chichester (2001).
> With regards,
> F. Pesarin
>
>
> At 07.21 11/04/02 -0500, you wrote:
> >Just convert the data to ranks and then run a regular 3-way anova.
> >
> >
> >----- Original Message -----
> >From: "Shirley Coleman"
> >To:
> >Sent: Thursday, April 11, 2002 3:42 AM
> >Subject: FW: non parametric problems
> >
> >
> >> Any ideas , please?
> >>
> >>
> >>
> >> I have been approached by a postgrad who has collected scored data from
a
> >> 3-way factorial experiment. There are only 9 classes and usually the
> >> replicates all have the same result. As you can imagine the residuals
are
> >> not normally distributed. Do you happen to know of a three-way
> >> non-parametric test that will let me also check the interactions
between
> >> the treatments?
> >>
> >
> >
> Prof. Fortunato Pesarin
> Department of Statistics
> University of Padova
> via Cesare Battisti, 241/243
> 35121 PADOVA - ITALY
> Tel. +39.049.8274143
> Fax +39.049.8274170
> email: [log in to unmask]
> Web page: www.stat.unipd.it/~pesarin
>
Phillip Good
http.ms//www.statistician.usa
I turn on my computer: therefore I am.
---------------------------------
Do You Yahoo!?
Yahoo! Tax Center - online filing with TurboTax
|