[R] confusing about contrasts concept [long]

[Peter Alspach]

Tian

It appears the attachment might not have worked so I'll embed Bill's
message at the end.

Peter Alspach


> -----Original Message-----
> From: r-help-bounces at stat.math.ethz.ch
> [mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of Peter Alspach
> Sent: Thursday, 17 August 2006 8:02 a.m.
> To: T Mu; R-Help
> Subject: Re: [R] confusing about contrasts concept
>
>
> Tian
>
> Bill Venables wrote an excellent explanation to the S list
> back in 1997.
> I saved it as a pdf file and attach it herewith ...
>
> Peter Alspach
>  
>
> > -----Original Message-----
> > From: r-help-bounces at stat.math.ethz.ch
>
> > [mailto:r-help-bounces at stat.math.ethz.ch] On Behalf Of T Mu
> > Sent: Thursday, 17 August 2006 3:23 a.m.
> > To: R-Help
> > Subject: [R] confusing about contrasts concept
> >
>
> > Hi all,
> >
>
> > Where can I find a thorough explanation of Contrasts and
>
> > Contrasts Matrices?
> > I read some help but still confused.
> >
>
> > Thank you,
> > Tian


Date sent:	Sat, 29 Mar 1997 17:23:31 +1030
From:		Bill Venables <wvenable at attunga.stats.adelaide.edu.au>
To:		Wei Qian <wwqian at venus.monsanto.com>
Copies to:	S-news at utstat.toronto.edu
Subject:	Re: test contrasts [long]

Wei Qian writes:

I am new to Splus, so please don't be offended if my question is too
naive.

We're all friends here, Wei. It's not that kind of group...mostly.

Does anyone know how to test contrasts in Splus? To be specific, suppose
I have a one-way layout experiment with 3 treatments, and I want to test
the contrasts of treatment 1 vs. treatment 2, treatment 1 vs. treatment
3, etc. In SAS, I could use the following:

 proc gim;
   class trt;
  model y = trt;
  contrasts"! vs. 2" 1-10;
  contrasts "2 vs. 3" 10-1;
run;

 One way I can think of is that to construct my design matrix and obtain
the contrast sum of squares through a series of matrix operations, but
is there any easy way to do it or any built-in function in Splus can do
it?

The answer is 'yes' but hardly anyone seems to know how to do it. The
explanation in the 'white book' for example, seems a little incomplete
to me and not quite adequate to settle the case you raise. (The
explanation in the yellow book is also inadequate, but I hope come July
we will have fixed that.) Since this is one of the most frequent
questions people ask me in direct email, too, let me try (again) to sort
it out in some detail.

A formula such as y ~ f, where f is a factor in principle generates a
single classification model in the form

*y_{ij} == mu + phi_i + e_{ij}

Write the design matrix in the form X = [1 Xf], where, assuming f has p
levels, Xf is the n x p dummy variable (ie binary) matrix corresponding
to the phi_i's. So in matrix terms the model is written as

*y = 1 mu + Xf phi + e

(a)	If you remove the intercept term, using y ~ f -1, then Xf is the
design matrix you get and the coefficients correspond to the class
means;
(b)	If the intercept term is left in, then the design matrix X does
not have full column rank, so an adjustment has to be made to Xf to make
it so.

The redundancy comes about because the columns of Xf add to the the
1-vector, that is
Xf l_p = l_n. So let C be any p x (p -1) matrix such that [1 C] is
nonsingular. It can easily be seen that Xc = [1 (Xf C)] will have full
column rank and that fitting the model using this model matrix is
equivalent to the original redundantly specified model. The matrix C is
called the *coding matrix* for the factor f.

The linear model is actually fitted in the form

*y = 1 mu + (Xf C) alpha + e

where alpha has (p-1) components, of course. In order to make sense of
the alpha's we need to relate them back to the phi's.

For any such C there will be a vector, v, such the v'C = 0' (using ' for
transpose). (In fact v spans the orthogonal complement to the column
space of C). Clearly phi and alpha are related by

*C alpha = phi

but since v'C = 0', it follows that an identification constraint
applies, namely v'phi = 0. By multiplying both sides by (C'C)^{-1} C',
it also follows that

*alpha =(C'C)^{-1}C'phi

which provides an interpretation for the alpha's in terms of the
(constrained) phi's. For example take the Helmert contrasts.

> contr.helmert(4)
[,1]	[,2]	[,3]
1	-1	-1	-1
2	1	-1	-1
3	0	2	-1
4	0	0	3

The constraint vector is clearly v= (1,1,1,1), since the columns add to
zero. In this case the columns are also mutually orthogonal, so the
matrix (C'C^{-l} C' (the generalized inverse of C) has a similar form
apart from a few row divisors:

>fractions(ginverse(contr.helmert(4)))
[,1]	[,2]	[,3]	[,4]
[1,]	-1/2	1/2	0	0
[2,]	-1/6	-1/6	1/3	0
[3,]	-1/12	-1/12	-1/12	1/4

(ginverse() will be available in S+4.0 and fractions(), now available in
the MASS2 library, simply displays numbers in fractional form so that
patterns are more obvious).

Thus the phi's are identified by requiring that they add to zero, and

*alpha_l = (phi_2 - phi_l )/2,
*alpha_2 = [phi_3 - (phi_l + phi_2)/2] / 3

&c. When the columns of C are not mutually orthogonal the story is not
quite so obvious, though. For example take a C matrix using contr.sdif
(available in the MASS2 library)

> contr.sdif(4)
[2-1]	[3-2]	[4-3]
1	-0.75	-0.5	-0.25
2	0.25	-0.5	-0.25
3	0.25	0.5	-0.25
4	0.25	0.5	0.75

The column sums are all zero so the identification constraint is still
the same, but they are not mutually orthogonal. The coefficients do have
an easy interpretation, though:

> fractions(ginverse(contr. sdif(4)))
[,1]	[,2]	[,3]	[,4]
[l,]	-1	1	0	0
[2,]	0	-1	1	0
[3,]	0	0	-1	1

Thus alpha_i = phi_(i+l) - phi_i, the "successive differences" of phi's,
or equally of class means.

Now (at last) returning to your case. In specifying the contrasts you
want to have as the coefficients, you are specifying not C, but
(C'C^{-l} C' (say M). So what you need to do is work backwards, noting
that C is also the ginverse of M:

> M <- rbind(c(l,-l,0), c(l, 0, -1))
>M
[,1]	[,2]	[,3]
[1,]	1	-1	0
[2,]	1	0	-1
> fractions(ginverse(M))
[,1]	[,2]
[1,]	1/3	1/3
[2,]	-2/3	1/3
[3,]	1/3	-2/3

This is the C matrix you need to specify as the coding matrix. Note that
(of course) the columns add to zero so that the implied constraint on
the phi's is that they sum to zero.

Since ginverse() is not yet published code, here is a quick and dirty
version that is adequate for this kind of case (but don't push it...)

ginv <- function(M)
*if(nrow(M) < ncol(M)) t(ginv(t(M)))
*else solve(crossprod(M), t(M))

So all you need to do is as follows:

> M <- rbind(c(l, -1, 0), c(l, 0, -1))
> Cmat <- ginv(M)
> dimnames(Cmat) <- list(NULL, c('l vs 2', '1 vs 3')) # optional
> contrasts(trt) <- Cmat
> fm <- aov(y ~ trt, .....)

The easy way to see the single degrees of freedom components is:

> summary .lm(fm)$coef

and the partitioned anova table can be shown using:

> summary(fm, split = list(trt = list('l vs 2' = 1, '1 vs 3'= 2)))

Finally you can see the complete phi-style representation of the fitted
model by

> dummy.coef(fm)

(Unfortunately dummy.coef does not generate an effective variance matrix
for the coefficients, otherwise contrasts could be directly tested after
the fitting process. This could be fixed, though.)

Note that the standard contrast functions contr.helmert, contr.poly and
contr.sum (as well as out contr.sdif) all impose the same identification
constraint on the phi's, i.e. they add to 1, but contr.treatment is
different as it corresponds to phi_l = 0. So the phi parameters mean
different things in this case, but the same principle applies. If the
rows of M all add to zero (making them "conventional contrasts") so will
the columns of C = M'(MM')^{-1}, (obviously).

Bill Venables.

Bill Venables, Head, Dept of Statistics,	Tel.: +61 8 8303 5418
University of Adelaide,			Fax.: +61 8 8303 3696
South AUSTRALIA. 5005.			Email:
Bill.Venables at adelaide.edu.au


______________________________________________________

The contents of this e-mail are privileged and/or confidenti...{{dropped}}