Re: GLM vs unbalanced designs

[Jason Stover]

On Sat, Nov 12, 2011 at 08:22:00AM +0000, John Darrington wrote:
> I've been reading a bit about two way anova with unbalanced designs.
> 
> I concluded that we should have been using what the literature calls 
> "effects coding" http://www.ats.ucla.edu/stat/mult_pkg/faq/general/effect.htm
> instead of "dummy coding" 
> http://www.ats.ucla.edu/stat/mult_pkg/faq/general/dummy.htm which is what we 
> have been using until now.
> 
> However I tried hacking up a quick change to use "effects coding", but 
> unfortunately it didn't improve anything.
> 
> Am I barking up the wrong tree here?

I just checked Neter, Wasserman and Kutner's "Applied Linear
Statistical Models", and I think you are right. The model PSPP uses
has different terms in its model by virtue of dummy coding. The
discrepancy could be fixed by adding the "effects coding" to
categoricals.c.

I would still keep dummy coding in categoricals.c as an option. It has
its place in other kinds of models.

-Jason