home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!gatech!paladin.american.edu!auvm!URIACC.BITNET!MARSH
- Message-ID: <STAT-L%92073011382348@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Thu, 30 Jul 1992 11:27:26 EDT
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: Marshall Feldman <MARSH@URIACC.BITNET>
- Subject: Re: interaction effects in regression models
- Lines: 29
-
- Three comments on the question of including insignificant variables
- that become significant in interactions:
-
- 1) I don't think one can make hard and fast rules, such as always include
- or never include the variable. The choice has to depend on the
- underlying theory (e.g., race by itself may be a significant
- predictor of income, but when race interacts with class and
- education, race by itself may become insignificant; this would
- be an important finding). So there is no hard and fast rule.
-
- 2) Generally, however, including both the variable and it interaction
- demonstrates that the interaction is not acting as a proxy for the
- variable.
-
- 3) Reviewers usually feel they have to say something intelligent or they
- are not doing their job. I am certain that many refereed articles
- are needlessly revised because of some reviewers' urge to seem
- to contribute to the process. In short, the best strategy for you
- might be to have a table with the results of three regressions
- (no interaction, interaction only, and both interaction and plain
- variable). Give the referee what s/he wants, but interpret the model
- you feel is most meaningful based on your understanding of the
- problem.
-
- Marsh Feldman
- Community Planning Phone: 401/792-2248
- 204 Rodman Hall FAX: 401/792-4395
- University of Rhode Island Internet: marsh@uriacc.uri.edu
- Kingston, RI 02881-0815
-