Re: interpreting significant interactions in Multiple Regression

Old Mac User wrote:
The following data came from a slightly
distorted 2-level factorial design. The fact
that the variables V1 and V2 are somewhat
correlated doesn't change the lesson to be
learned from this.

In the following I set the centering constant
for V2 to -24. The centering constant for
V1 has been set to -5, -6, -7, and -8.

Notice that the t-ratio for V2 varies with the
centering constant for V1.

Yes, if all terms are included in the model
then the predicted values are the same
no matter what centering constants are
used. But if the choice of constant (using
no centering constant at all is the same
as using zero) then the t-ratio for V2 may
be so low as to cause us to eliminate
that term from the model. In which case
the game changes dramatically.

[... data snipped ...]

I agree that if a model contains terms x1, x2, and x1*x2 then adding
a constant to x1 will change the weight on x2, and adding a constant
to x2 will change the weight on x1. In fact, constants can always be
chosen to make the weights on both x1 and x2 zero, without changing
the weight on x1*x2.

But that's not the issue here. The OP's model has BA, NATU, MAST,
BA*NATU, and BA*MAST. The point I was trying to make is that
centering NATU and/or MAST will not change the weight (or the t-test
of the weight) on NATU, MAST, BA*NATU, or BA*MAST. Only the constant
and the weight on BA (and their associated t-tests) will change.

.