Date:         Sat, 26 Aug 2000 21:22:17 +0200
Reply-To:     Dirk Enzmann <enzmann@KFN.UNI-HANNOVER.DE>
Sender:       "SPSSX(r) Discussion" <SPSSX-L@LISTSERV.UGA.EDU>
From:         Dirk Enzmann <enzmann@KFN.UNI-HANNOVER.DE>
Organization: Kriminologisches Forschungsinstitut Hannover e.V.
Subject:      Re: testing for interactions in regression
Content-Type: multipart/alternative;

Marcel,

Before computing the interaction term, in most cases the variables are
centered (i.e. transformed to deviation scores). The difference between
centered variables and z-scores (standardized variables) is that
centering preserves the original variance of the variables. If you do
not center your variables and the variables are scaled positively with a
real zero point (i.e. from 0 up), the regression coefficient for each
constituent variable estimates the effect of that variable when all
other components are zero. If you center the variables before computing
the interaction term, in the presence of a significant interaction
effect a main effect can be interpreted as the effect of a variable at
the average observed score on the moderator variable. See also: Finney,
J.W., Mitchell, R.E., Cronkite, R.C. & Moos, R.H. (1984). Methodological
issues in estimating main and interaction effects: Examples from
coping/social support and stress field. Journal of Health and Social
Behavior, 25, 85-98.

A SPSS macro to compute centered variables with listwise deletion can be
obtained from  http://www.kfn.de/softwareenzmann.html

Dirk

Marcel van Aken wrote:

> I want to test the interaction of two continuous variables (x1 and x2) in
> predicting a third (y). I want to do this by entering the two main
> variables as first step in a hierarchical regression, and then adding the
> product (x1*x2) as a second step. I noticed that there are huge differences
> in results, depending on whether or not I standardize the variables before
> computing the interaction term.
> I checked several empirical examples of this procedure, but either no-one
> mentions the fact that their variables are standardized, or no-one
> standardizes them.
> What should I do? Could someone point me to a clear text about this issue
> (not a global book on regression, these are not always clear about this
> point either).

*************************************************
Dr. Dirk Enzmann
Criminological Research Institute of Lower Saxony
Luetzerodestr. 9
D-30161 Hannover
Germany

phone: +49-511-348.36.32
fax:   +49-511-348.36.10
email: ENZMANN@KFN.uni-hannover.de

http://www.kfn.de
*************************************************


[text/html]