moderating relationship and regression

[Marnie LaNoue]

Hello List!

I am looking for some analytic advice for the following problem: it is
hypothesized that an X --> Y relationship is moderated by 'Z'. X is a
grouping variable and Y is continuously distributed. Z is also coninuously
distributed. (the actual research hypothesis goes something like this: in
all groups except one, the moderating variable will be negatively related
to the outcome. In the one group of interest, the slope of the line between
the moderator and the outcome will be near zero ).

Any advice on how to analyze this problem? If I choose to go with a
regression framework, is it simply a matter of creating prodcut
(interaction) terms and regressing the outcome on them?

Additional concern is how to compute power for this design.

Any insight would be appreciated!

Marnie

[Hector Maletta]

The number of groups (categories of X) is large or small? And do you know in
advance which is the one group of interest? If the number of groups is small
and you can know in advance which one is the group of interest, then you can
probably approach the problem in a very simple way, examining the ZY
relationship group by group to check whether the model predictions are
fulfilled by the data (i.e. whether the relationship is significantly
inverse, with a negative beta coefficient for Z in all but one of the groups
and beta not significantly different from zero in the one group of
interest). If not, you may need other approaches.

Hector

-----Mensaje original-----
De: SPSSX(r) Discussion [mailto:[hidden email]] En nombre de
Marnie LaNoue
Enviado el: 09 October 2006 14:23
Para: [hidden email]
Asunto: moderating relationship and regression

Hello List!

I am looking for some analytic advice for the following problem: it is
hypothesized that an X --> Y relationship is moderated by 'Z'. X is a
grouping variable and Y is continuously distributed. Z is also coninuously
distributed. (the actual research hypothesis goes something like this: in
all groups except one, the moderating variable will be negatively related
to the outcome. In the one group of interest, the slope of the line between
the moderator and the outcome will be near zero ).

Any advice on how to analyze this problem? If I choose to go with a
regression framework, is it simply a matter of creating prodcut
(interaction) terms and regressing the outcome on them?

Additional concern is how to compute power for this design.

Any insight would be appreciated!

Marnie

[Maguin, Eugene]

Marnie,

Unless I misunderstand your question, I think you can form two supergroups,
one of which consists of all groups but the one where the regression of Y on
Z is theorized to be negative and the other which consists of only that
group, and then regress Y on X, Z and X*Z. X is simply a dichotomous
variable, coded 0 or 1, indicating group membership and X*Z is just that--X
times Z. So yes, is just a matter of creating the product term.

> Additional concern is how to compute power for this design.

One way to get power is through a power calculation program, some of which
are available on the web, or you can work your way through a power
calculation book, for example, Jacob Cohen. Spss regression won't calculate
power. But, spss GLM will and I think that you can set up your regression as
a one between factor model with a covariate and the factor by covariate
interaction specified in the design subcommand and then ask for the power
computation in the print subcommand.

If I am incorrect about this, I hope that someone will jump in and correct
my error.

Gene Maguin