home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!WATSON.BITNET!FOLGER
- Return-Path: <Folger@OLIVAW.watson.ibm.com>
- X-External-Networks: yes
- Message-ID: <9301260012.AA0498@OLIVAW.watson.ibm.com>
- Newsgroups: bit.listserv.stat-l
- Date: Mon, 25 Jan 1993 18:22:45 EST
- Reply-To: "Davis A. Foulger (914) 945-2077 (t. 862-2077)"
- <Foulger@WATSON.BITNET>
- Sender: STATISTICAL CONSULTING <STAT-L@MCGILL1.BITNET>
- From: FOLGER@WATSON.BITNET
- Subject: interaction effects
- Lines: 43
-
- I certainly agree with Don Ploch that in testing an interaction one should
- test a multiplicative one, but still note that, as long as one is doing
- regression and has the freedom to shape the interaction, one should be aware
- of the form the hypothesized interaction is expected to take.
-
- There are a variety of multiplicative interaction shapes that can be tested,
- including AND, OR, XOR, NAND, NOR, NXOR, nonzero positive, and various
- combinations thereof. All are equivalent in terms of the total variance
- accounted for when both interactions and main effects are tested. Each
- shape represents a very different interpretation of the data, however, such
- that selection of the right shape (e.g. the hypothesized interaction) will
- make the results of the analysis easier to interpret; selection of the wrong
- shape may make misinterpretation likely.
-
- Consider, for instance, the classic interaction hypothesis that one effect
- happens only in the presence of condition 1 of variable X and condition 1 of
- variable Y (e.g. not in the presence of condition 0 of either variable --
- an AND interaction if you will). Assuming these results obtain, the
- standard XOR interaction used in an analysis of variance (with the main
- effects in the two x two condition coded as -1 and 1; condition 0 and
- condition 1, respectively) will show two significant main effects and a
- significant interaction; this despite the lack of any real main effect.
- We can, however, easily code the interaction to the hypothesized shape by
- coding the conditions as 0 and 1 (condition 0 and condition 1, respectively;
- the AND interaction). Again assuming that the results obtain as
- hypothesized, only the interaction will be significant; neither of the main
- effects will be.
-
- These are only 2 of a variety of possible hypothesized shapes (although they
- are the principle shapes for the simple bivariate case). It should be
- obvious, however, from my description, that, at least for the hypothesis
- given, the second shape above is more likely to be correctly interpreted
- than the first. This problem of interpretability is one of the reasons I am
- very much against the routine use of ANOVA in testing interaction effects.
- I am sure that many a main effect has been incorrectly reported as
- significant because an AND interaction was tested by ANOVA's standard XOR
- interaction shape.
-
- Davis
-
- Snailmail..........................Davis A. Foulger
- Internet: FOLGER@WATSON.IBM.COM IBM T.J. Watson Research Center
- Prodigy: XFRR20A P O Box 218, Yorktown Ht, NY 10598
-