R help - Oct 2010 - creating 'all' sum contrasts

OK, my last question didn't get any replies so I am going to try and ask a
different way.

When I generate contrasts with contr.sum() for a 3 level categorical variable I
get the 2 orthogonal contrasts:
> contr.sum( c(1,2,3) )
  [,1] [,2]
1    1    0
2    0    1
3   -1   -1

This provides the contrasts <1-3> and <2-3> as expected.  But I also
want it to create <1-2> (i.e. <1-3> - <2-3>).  So in general I
want all possible orthogonal contrasts - think of it as the contrasts for all
pairwise comparisons between the levels.

Are there are any options for contrast() or other functions/libraries that will
allow me to do this automatically?  I could go through and create new columns
but I am using this for complex multi-factor experiments with varying levels per
factor and fitting the models within other functions (e.g. regsubsets()) so an
automatic solution using what is already available would be far preferable.


Michael Hopkins
Algorithm and Statistical Modelling Expert
 
Upstream
23 Old Bond Street
London
W1S 4PZ


	[[alternative HTML version deleted]]

G'day Michael,

On Fri, 15 Oct 2010 12:09:07 +0100
Michael Hopkins <hopkins at upstreamsystems.com> wrote:
> OK, my last question didn't get any replies so I am going to try and
> ask a different way.
> 
> When I generate contrasts with contr.sum() for a 3 level categorical
> variable I get the 2 orthogonal contrasts:
> 
> > contr.sum( c(1,2,3) )
>   [,1] [,2]
> 1    1    0
> 2    0    1
> 3   -1   -1
These two contrasts are *not* orthogonal.
> This provides the contrasts <1-3> and <2-3> as expected.  But I
also
> want it to create <1-2> (i.e. <1-3> - <2-3>).  So in
general I want
> all possible orthogonal contrasts - think of it as the contrasts for
> all pairwise comparisons between the levels.
You have to decide what you want.  The contrasts for all pairwise
comparaisons between the levels or all possible orthogonal contrasts?

The latter is actually not well defined.  For a factor with p levels,
there would be p orthogonal contrasts, which are only identifiable up to
rotation, hence infinitely many such sets. But there are p(p-1)
pairwise comparisons. So unless p=2, yo have to decide what you want....
> Are there are any options for contrast() or other functions/libraries
> that will allow me to do this automatically?
Look at package multcomp, in particular functions glht and mcp, these
might help.

Cheers,

	Berwin

========================== Full address ===========================Berwin A
Turlach                      Tel.: +61 (8) 6488 3338 (secr)
School of Maths and Stats (M019)            +61 (8) 6488 3383 (self)
The University of Western Australia   FAX : +61 (8) 6488 1028
35 Stirling Highway                   
Crawley WA 6009                e-mail: berwin at maths.uwa.edu.au
Australia                        http://www.maths.uwa.edu.au/~berwin