How to select correlated independents?
Frank
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.
mcap
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.
Marc
Rich Ulrich
wrote:
Frank wants to use as many predictors as possible. Marc warns
of compromises.
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.
--
Rich Ulrich
Frank
Frank