On Sat, 7 Nov 2009 06:42:02 -0800 (PST), mcap <firstname.lastname@example.org>
>On Nov 6, 4:53�pm, Frank <danbomingzhi1...@sina.com> wrote:
>> Any help will be appreciated greatly. I have one dependent y, and 4
>> indepent varaibles x1 x2 x3 x4 are correlated, and 3 indep z1 z2 z3
>> are correlated, 4 indep w1 w2 w3 w4 are also correlated.
>> I would like to regress y on all 11 predictors to build one model, and
>> include them as many as possible.
>Run collinearity diagnostics and don't forget to look for plausible
>effect modifiers. In the end, you may have to make this decision on a
>clinical basis, not a statistical basis. You could also use factor
>analysis or other methods. No matter what you do, limitations and
>compromises are coming.
Frank wants to use as many predictors as possible. Marc warns
Marc could be jumping the gun. Frank apparently expects a
problem, but he did not establish that that is any *real* problem.
What happens when the equation uses them all? Diagnositics?
And then? What N is available to work with, to establish what
sort of R^2?
If the W's, X's and Z's each represent a sensible "latent variable",
and those latent variables are expected to carry the prediction,
then there is also no problem -- The approach *should* be
shortened to one that creates the 3 composite variables and
then uses those three for the prediction.