R help - Aug 2010 - Specify contrasts in R

Dear Listers,

Could anyone advise if this is the 'correct' contrast matrix to use to
make contrasts between one treatment and the mean of 2 other treatments (in this
case, mid-parent values to the F1 and F2 from a line cross experiment)?
Desired contrasts are
1) P1 versus P2
2) average of P1 and P2 versus F1
3) average of P1 and P2 versus F2
levels(MyData$Type)
"P1", "F1", "F2","P2"
(I have renamed these A, B, C, D in the actual data so they appear in the order
above)
Contrast matrix:
contr_mat<-cbind(
 c(1,0,0,-1),
c(-1,2,0,-1),
c(-1,0,2,-1))

contrasts(MyData$Type)<-contr_mat

Model:
mod<-lm(Trait~Type,data=MyData)



Also, I would ideally like to run this in lme with a random effect for Block

mod2<-lme(Trait~Type,random=~1|Block,data=MyData)



Each treatment appears once in each block (ie balanced). Is there any problem
using this contrast set-up in a mixed model? I am not sure how the
'block' effect would be interpreted for the mid-parent values, or indeed
if it is sensible to include it as a random effect.



Thank you in advance for your help,



Rebecca Ross

University of Oxford









	[[alternative HTML version deleted]]

On Aug 2, 2010, at 3:53 PM, Rebecca Ross wrote:
> Could anyone advise if this is the 'correct' contrast matrix to use
to make contrasts between one treatment and the mean of 2 other treatments (in
this case, mid-parent values to the F1 and F2 from a line cross experiment)?
> Desired contrasts are
> 1) P1 versus P2
> 2) average of P1 and P2 versus F1
> 3) average of P1 and P2 versus F2
> levels(MyData$Type)
> "P1", "F1", "F2","P2"
> (I have renamed these A, B, C, D in the actual data so they appear in the
order above)
> Contrast matrix:
> contr_mat<-cbind(
> c(1,0,0,-1),
> c(-1,2,0,-1),
> c(-1,0,2,-1))
> 
> contrasts(MyData$Type)<-contr_mat
If you have played with this for some time, you develop a strong aversion to
non-orthogonal contrast matrices...

You have to be very careful to notice that contrast matrices map FROM parameters
TO mean values, whereas what we normally call contrasts map in the other
direction.

In general, to get a set of parameters that has a particular contrast
parametrization, you have to do a (generalized) matrix inverse. In the present
case,
> library(MASS)
> zapsmall(ginv(rbind(c(1,0,0,-1),c(-.5,1,0,-.5),c(-.5,0,1,-.5))))
     [,1]  [,2]  [,3]
[1,]  0.5 -0.25 -0.25
[2,]  0.0  0.75 -0.25
[3,]  0.0 -0.25  0.75
[4,] -0.5 -0.25 -0.25

which looks pretty whacked when you first look at it, but to wit:
> CM <- zapsmall(ginv(rbind(c(1,0,0,-1),c(-.5,1,0,-.5),c(-.5,0,1,-.5))))
> solve(cbind(1,CM))
      [,1] [,2] [,3]  [,4]
[1,]  0.25 0.25 0.25  0.25
[2,]  1.00 0.00 0.00 -1.00
[3,] -0.50 1.00 0.00 -0.50
[4,] -0.50 0.00 1.00 -0.50

So, no, you guessed incorrectly.


-- 
Peter Dalgaard
Center for Statistics, Copenhagen Business School
Solbjerg Plads 3, 2000 Frederiksberg, Denmark
Phone: (+45)38153501
Email: pd.mes at cbs.dk  Priv: PDalgd at gmail.com