R help - Aug 2004 - drop1 with contr.treatment

Dear R Core Team

I've a proposal to improve drop1(). The function should change the
contrast from the default ("treatment") to "sum". If you fit
a
model with an interaction (which ist not signifikant) and you
display the main effect with

drop1(   , scope = .~., test = "F")

If you remove the interaction, then everything's okay. There is
no way to fit a model, let an non signifikant interaction in
the model and get interpretable main effects with the
"treatment" contrast.

One solution would be to change automatically the contrast to
the "sum" contrast.
Another possibility is to produce a warning when you use drop1
with the scope argument to get main effects in the presence of 
an interaction (even not signifikant) with "treatment"
contrast. 

An example:

library(MASS)
##- Data "genotype"
names(genotype)
##- > [1] "Litter" "Mother" "Wt"

##- to be sure the contrasts are "treatment"
options(contrasts = c("contr.treatment", "contr.poly" ))

##- model with interaction
gen.int <- aov(Wt ~ Litter*Mother, data = genotype)
drop1(gen.int, scope = .~., test = "F")
##- Model:
##- Wt ~ Litter * Mother
##-               Df Sum of Sq    RSS    AIC F value   Pr(F)  
##- <none>                     2440.8  257.0                  
##- Litter         3     591.7 3032.5  264.3  3.6362 0.01968 *
##- Mother         3     582.3 3023.1  264.1  3.5782 0.02099 *
##- Litter:Mother  9     824.1 3264.9  256.8  1.6881 0.12005  
##- ---
##- Signif. codes:  0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 `
' 1

##- the interaction is not signifikant, Litter is signifikant,
##- but is that really true?

##- main effect model
gen.main <- aov(Wt ~ Litter + Mother, data = genotype)
drop1(gen.main, scope = .~., test = "F")
##- Model:
##- Wt ~ Litter + Mother
##-        Df Sum of Sq    RSS    AIC F value   Pr(F)   
##- <none>              3264.9  256.8                   
##- Litter  3      63.6 3328.5  252.0  0.3508 0.78870   
##- Mother  3     775.1 4040.0  263.8  4.2732 0.00886 **
##- ---
##- Signif. codes:  0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 `
' 1

## Dramatic change of the sum of squares for the factor Litter

## With contrast = "sum" the whole thing looks better
options(contrasts = c("contr.sum", "contr.poly" ))

##- model with interaction
gen.int1 <- aov(Wt ~ Litter*Mother, data = genotype)
drop1(gen.int1, scope = .~., test = "F")
##- Model:
##- Wt ~ Litter * Mother
##-               Df Sum of Sq    RSS    AIC F value   Pr(F)  
##- <none>                     2440.8  257.0                  
##- Litter         3      27.7 2468.5  251.7  0.1700 0.91612  
##- Mother         3     671.7 3112.6  265.9  4.1282 0.01142 *
##- Litter:Mother  9     824.1 3264.9  256.8  1.6881 0.12005  
##- ---
##- Signif. codes:  0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 `
' 1

##- the interaction is not signifikant, Litter is NOT! signifikant.

gen.main1 <- aov(Wt ~ Litter + Mother, data = genotype)
drop1(gen.main1, scope = .~., test = "F")
##- Wt ~ Litter + Mother
##-        Df Sum of Sq    RSS    AIC F value   Pr(F)   
##- <none>              3264.9  256.8                   
##- Litter  3      63.6 3328.5  252.0  0.3508 0.78870   
##- Mother  3     775.1 4040.0  263.8  4.2732 0.00886 **
##- ---
##- Signif. codes:  0 `***' 0.001 `**' 0.01 `*' 0.05 `.' 0.1 `
' 1

##- Litter stays not signifikant.


Best regards

Christoph Buser

-- 
Christoph Buser <buser at stat.math.ethz.ch>
Seminar fuer Statistik, LEO D6
ETH (Federal Inst. Technology)	8092 Zurich	 SWITZERLAND
phone: x-41-1-632-3505		fax: 632-1228
http://stat.ethz.ch/~buser/

Since this was addressed to R-Core (please do read the posting guide!) 
discussion will happen elsewhere.

R-help is not the right place for suggestions of changes to R: R-devel may
be.  As with bugs, please ensure you understand why the existing
behaviour is not both correct and intentional.  In a case like this, see
what the White Book says it does (as referred to on the help page) and
check it against S if you can.

On Fri, 20 Aug 2004, Christoph Buser wrote:
> Dear R Core Team
> 
> I've a proposal to improve drop1(). The function should change the
> contrast from the default ("treatment") to "sum". If
you fit a
> model with an interaction (which ist not signifikant) and you
> display the main effect with
> 
> drop1(   , scope = .~., test = "F")
> 
> If you remove the interaction, then everything's okay. There is
> no way to fit a model, let an non signifikant interaction in
> the model and get interpretable main effects with the
> "treatment" contrast.
> 
> One solution would be to change automatically the contrast to
> the "sum" contrast.
> Another possibility is to produce a warning when you use drop1
> with the scope argument to get main effects in the presence of 
> an interaction (even not signifikant) with "treatment"
> contrast. 
[...]

-- 
Brian D. Ripley,                  ripley at stats.ox.ac.uk
Professor of Applied Statistics,  http://www.stats.ox.ac.uk/~ripley/
University of Oxford,             Tel:  +44 1865 272861 (self)
1 South Parks Road,                     +44 1865 272866 (PA)
Oxford OX1 3TG, UK                Fax:  +44 1865 272595