I have a couple questions:
(1) Is VIF high without the interaction
terms?
(2) What is the mean correlation among
your predictors and
the range correlations? If these are
high, adding the interactions
terms will just worsen the
situation.
(3) What is the squared multiple
correlation (SMC) of each predictor
with the other predictors? If these are
high, adding the
interaction terms will worsen the
situation.
Answers to questions (2) and (3) can be
obtained via the reliability
procedure and treating your predictors as
though it were a scale
(of course, it's not but what you want are
summary info for the
pariwise correlations and the
SMCs).
If there are highly correlated predictors,
this is an indication that
your predictors are providing redundant
information. Either select
the best predictors or combine them in a way
that would make the
combined variable less correlated with the
other predictors.
By the way, I'm assuming you're using
regression because you
have data from a nonexperimental research
study. In this type of
situation the correlations between
predictors may be real or spurious,
that is, due to a third variable(s). You
need to know your variables,
how they interrelate to the variables in your
equation as well as
variables not in your equation (i.e., third
variables that will lead
to model misspecification).
-Mike Palij
New York University
----- Original Message -----
Sent: Thursday, March 10, 2011 10:15
AM
Subject: Multicollinearity after
centering: interaction variable
Dear all,
I am performing a linear regression
analysis with multiple independent variables and some interaction variables.
Of course the interaction variables gave multicollinearity problems so I
centered the variables around the mean and computed the interaction variables
with the centered variables and performed another regression analysis.
Unfortunatly the results still showed very very high VIF values indicating
multicollinearity. Now I am kind of lost. What should I do to include the
interaction terms but not have multicollinearity problems? By the way, for
computing the interaction terms I use one dummy variable and one continous
variable.
Thanks for any help.
Locked
