ANOVA procedure on the sufficient statistics

You can apply the trick that Prof. Dalgaard recently posted in response to a
similar question (for one-way ANOVA).  For each cell, generate data as:

y <- cell.mean + cell.sd * scale(rnorm(cell.count))

Then generate the data frame to feed to aov.

 

That will generate only approximately the same means and MSE, however.

Larsen's procedure generates a weighted data set that gives the same 
ANOVA table
as the raw data, but requires that the weight= argument be used in aov().
(weights are not supported in all related functions)

On Thu, 19 Feb 2004 08:45:01 -0500, Michael Friendly
<friendly at yorku.ca> wrote :

That will generate only approximately the same means and MSE, however.

Larsen's procedure generates a weighted data set that gives the same 
ANOVA table
as the raw data, but requires that the weight= argument be used in aov().
(weights are not supported in all related functions)

Are you sure about that?  I think the call to "scale()" makes them
come out identical.

In fact, the call to rnorm() is unnecessary; it would work just as
well with 

y <- cell.mean + cell.sd * scale(1:cell.count)

(unless I'm missing something...)