R help - Nov 2006 - Interaction and factor ':'

_______________________________________________________________________________________


Hello,

Not sure if this counts as a bug or not, but I just noticed in R 2.4.0
that : and interaction are not quite equivalent. For example:
> x <- factor(letters[1:4])
> y <- factor(letters[11:14])
> x:y
[1] a:k b:l c:m d:n
Levels: a:k a:l a:m a:n b:k b:l b:m b:n c:k c:l c:m c:n d:k d:l d:m d:n
> interaction(x, y, sep=":")
[1] a:k b:l c:m d:n
Levels: a:k b:k c:k d:k a:l b:l c:l d:l a:m b:m c:m d:m a:n b:n c:n d:n

The ordering of the levels is different between the two, although
?interaction says


    f:g is the same as interaction(f,g, sep=":") when f and g are
factors.

I came across this when fitting multinom models, and the column order of
the predicted probabilities shifted around depending on whether I used
interaction or : on the LHS of the formula.


-- 
Hong Ooi
Senior Research Analyst, IAG Limited
388 George St, Sydney NSW 2000
+61 (2) 9292 1566

_______________________________________________________________________________________

The information transmitted in this message and its attachme...{{dropped}}

On Thu, 16 Nov 2006, Hong Ooi wrote:
> Not sure if this counts as a bug or not, but I just noticed in R 2.4.0
> that : and interaction are not quite equivalent. For example:
Well, that depends on what you mean by 'equivalent', a word that is not 
used in the help you quote but is used in help(":").  If you view a 
factor as an enumeration type, they are equivalent.

This difference seems to be long-standing (it was there in R 2.0.0), and 
I'd say it was unclear documentation which I will amend.

(As a historical note, interaction() behaves the same as S, whereas ':'
in
S does not have this alternative meaning but had been that way in R for 
many years.)
>
>> x <- factor(letters[1:4])
>> y <- factor(letters[11:14])
>
>> x:y
> [1] a:k b:l c:m d:n
> Levels: a:k a:l a:m a:n b:k b:l b:m b:n c:k c:l c:m c:n d:k d:l d:m d:n
>
>> interaction(x, y, sep=":")
> [1] a:k b:l c:m d:n
> Levels: a:k b:k c:k d:k a:l b:l c:l d:l a:m b:m c:m d:m a:n b:n c:n d:n
>
> The ordering of the levels is different between the two, although
> ?interaction says
>
>
>    f:g is the same as interaction(f,g, sep=":") when f and g are
> factors.
>
> I came across this when fitting multinom models, and the column order of
> the predicted probabilities shifted around depending on whether I used
> interaction or : on the LHS of the formula.
Which seems fine, as it is the labels and not the order which mattered.

-- 
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