home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!uvaarpa!darwin.sura.net!paladin.american.edu!auvm!UTKVX.BITNET!PLOCH
- Original_To: BITNET%"stat-l@mcgill1"
- Message-ID: <STAT-L%92073111474684@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Fri, 31 Jul 1992 11:44:00 EDT
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: PLOCH@UTKVX.BITNET
- Subject: RE: interaction effect
- Lines: 26
-
- One needs to retain lower order terms in the equation, because the
- presence of an interaction term means that the equation can be re-
- written as a multiplicative equation using linear transforms of
- the main effects.
- Y=b0 + b1X1 + b2X2 + b3X1X2
- can always be factored to
- Y=K + (a1 + a2X1) (a3 + a4X2)
- There are an infinite number of solutions since one must define one
- of the a(i) to get a solution. If b1 or b2 are set to zero, the
- solution is easier since a1 or a3 must be set to zero. The point is
- that even if set to zero, X1 and X2 continue to have effects. One
- might determine the relative size of these effects by solving for
- ln(Y-K)=c0 + c1(a1 + a2X1) + c2(a3 + a4X2)
-
- The arguement extends to higher orders of interaction and to
- polynomials. It does not extend to categorical variables.
-
- Retaining lower order terms that are not significant makes sense.
- More to the point is that the presence of interaction should alert
- the analyst to try non-linear models.
-
- Don Ploch
- (BITNET: Ploch@UTKVX)
- (INTERNET: PLOCH@UTKVX.UTK.EDU)
- Sociology, University of Tennessee, Knoxville TN 37996
- (615)974-7022
-