Re: Statistica anova

Bob Wheeler (bwheeler@ssnet.com)
Wed, 7 Sep 94 11:05:50 EDT


>
> A historical note might be in order for this discussion. Fisher
> hated ANOVA and frequently warned of its dangers. He wrote that
> ANOVA was nothing more than a convenient way of arranging the
> arithmetic. If Fisher and his colleagues had access to modern
> computing, specifically something that could do, say, a QR
> decomposion, ANOVA wouldn't have ever happened. We would all be
> specifying our contrasts and testing them in the context of the
> general linear model instead of running some canned ANOVA
> program. The dangers, which Fisher foresaw, are that concepts
> like "main effects" and "interaction" (as Gould has clearly noted
> in this discussion) are slippery outside the context of balanced
> designs with no empty cells. I'm amazed that we continue to
> confuse ourselves with these slippery concepts. I think we would
> be better off if all packages got rid of their anova programs.
> Now that we have computers we no longer need the convenient way
> of arranging the arithmetic.
>
> Gary.McClelland@Colorado.edu
> Psycholgy, Univ of Colorado
>

These are very important points.

There are a number of statistical things
that have outlived their utility. A master, such as Fisher, can
stamp a methodology so firmly into the minds of his followers that
it becomes almost impossible to replace it when it becomes obsolescent.

The advent of the computer and widely used statistical packages is
having the same effect that the advent of printing had on language --
the freezing of methods of expression. I can think of advantages
for language, but off the top of my head I cannot think of any for
statistics.

It is difficult to write a commercial computer program that departs
from materials that appear in the standard textbooks. If one does,
reviewers are sure to denigrate the program, and readers of the reviews
will look elsewhere. I believe Paul Velleman observed someplace on this
thread or another that there were better ways than stepwise regression
to deal with the problem, and that his Data Desk offers an alternative.
I think his observations were not given due weight.

In my case, experimental design for engineers, the standard textbooks
present a considerable problem. They are structured for statisticians,
and many of the methodologies depend on a world view alien to
engineers. Fortunately statistics is rich enough that one can choose
paths less well traveled -- criticism for not employing standard methodologies

is to be expected.

Bob Wheeler, ECHIP, Inc.