[R] paired t-test vs pairwise t-test

[Marc Schwartz]

On Thu, 2004-08-19 at 14:42, Liaw, Andy wrote:
> > From: Duncan Murdoch
> > 
> > On Thu, 19 Aug 2004 13:42:21 -0300 (ADT), Rolf Turner
> > <rolf at math.unb.ca> wrote :
> > 
> > >
> > >You wrote:
> > >
> > >> What's the difference between t.test(x, y) and 
> > pairwise.t.test()? Is
> > >> it just that the former takes two vectors, whereas the 
> > latter takes a
> > >> vector and a factor?
> > >
> > >	No.  The pairwise.t.test() function (according to the help
> > >	file) does a multiplicity of t-tests, on more than two
> > >	samples, adjusting the p-value to compensate for the
> > >	multiplicity by various methods.
> > >
> > >	IMHO the name of this function is bad, because to me it
> > >	suggests doing ***paired*** t-tests, which would trip up the
> > >	naive user, who probably wouldn't notice or would ignore the
> > >	"t tests with pooled SD" message in the output.  As one of
> > >	the Ripley fortunes says ``It really is hard to anticipate
> > >	just how silly users can be.''  But why go out of the way to
> > >	give them a chance to be silly?
> > 
> > And Jack wrote:
> > 
> > >But the documentation, which I valiantly tried to make sense 
> > of BEFORE 
> > >asking my stupid question, is not clear enough for this 
> > particular idiot. 
> > >Might I suggest that the documentation be altered? It could 
> > use an example 
> > >(as in, real-life applied statistical problem) of when 
> > pairwise.t.test() 
> > >ought to be used, and why t.test(paired=TRUE) would be 
> > inappropriate in that 
> > >context; it could also use a reference to some published 
> > paper, website or 
> > >some such that explains the rationale and correct procedure 
> > for using this 
> > >test.
> > 
> > I think it's unlikely that we would rename the function; it's been
> > around a while with its current name so that's a bad idea.  On the
> > other hand, clearer documentation is always a plus:  why not submit
> > some?
> 
> I guess this is sort of related to the thread on whether R is good for
> non-statisticians...  The help pages in R are sort of like *nix man pages.
> They give the technical information about the topic, but not necessarily the
> background.  E.g., the man page for `chmod' does not explain file
> permissions in detail: the user is expected to learn that elsewhere.
> 
> Perhaps other stat packages do it differently?  Does SPSS manuals detail
> what its t-test procedure does, including which t-test(s) it does and when
> it's appropriate?  That might make it easier on users, but I still think the
> users should learn the appropriate use of statistical procedures
> elsewhere...
> 
> Best,
> Andy


Andy,

I don't know about SPSS, but SAS' documentation is available online at:

http://support.sas.com/91doc/docMainpage.jsp

The documentation specifically for PROC TTEST is at:

http://support.sas.com/91doc/getDoc/statug.hlp/ttest_index.htm

and the documentation for PROC MULTTEST is at:

http://support.sas.com/91doc/getDoc/statug.hlp/multtest_index.htm

Of course, to go along with the standard SAS documentation, there is the
line of "Books by Users", which parallels in a fashion, the increasing
number of books on R, authored by members of this community.

Best regards,

Marc