MCA in R

Dear Kimmo,

-----Original Message-----
From: r-help-bounces at r-project.org [mailto:r-help-bounces at r-project.org]

On

Behalf Of K. Elo
Sent: June-13-08 1:43 AM
To: r-help at r-project.org
Subject: Re: [R] MCA in R

Dear John,

thanks for Your quick reply.

John Fox wrote:
Dear Kimmo,

MCA is a rather old name (introduced, I think, in the 1960s by
Songuist and Morgan in the OSIRIS package) for a linear model
consisting entirely of factors and with only additive effects --
i.e., an ANOVA model will no interactions.

It is true, that MCA is an old name, but the technique itself is still
robust, I think. The problem I am facing is that I have a research
project where I try to find out which factors affect measured knowledge
of a specific issue. As predictors I have formal education, interest,
gender and consumption of different medias (TV, newspapers etc.). Now,
these are correlated predictors and running e.g. a simple anova
(anova(lm(...)) as You suggested) won't - if I have understood correctly
- consider the problem of correlated predictors. MCA would do this.

That's because anova() calculates sequential ("type-I") sums of squares; if
you use the Anova() function in the car package, for example, you'll get
so-called type-II sums of squares -- for each factor after the others. You
could also more tediously do these tests directly using the anova()
function, by contrasting alternative models: the full model and the model
deleting each factor in turn.

A colleague of mine has run anova and MCA in SPSS and the results differ
significantly.

Yes, see above.

 Because I am more familiar with R, I just hoped that this
marvelous statistical package could handle MCA, too :)

Typically, the results of
an MCA are reported using "adjusted means." You could compute these
manually, or via the effects package.

Well, I am interested in the eta and beta values, too.

Aren't the eta values just the square-roots of the R^2's from the individual
one-way ANOVAs? I don't remember how the betas are defined, but do recall
that they are a peculiar attempt to define standardized partial regression
coefficients for factors that combine all of the levels.

I have tried to
use the effects package but my attempts with all.effects resulted in
errors. I have to figure out what's going wrong here :)

If you tell me what you did, ideally including an example that I can
reproduce, I can probably tell you what's wrong.

Regards,
John

Kind regards,
Kimmo Elo

--
University of Turku, Finland
Dep. of political science