home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!europa.asd.contel.com!paladin.american.edu!auvm!REED.EDU!JONES
- Message-ID: <m0mDdgz-0000iiC@deneb.reed.edu>
- Newsgroups: bit.listserv.stat-l
- Date: Thu, 30 Jul 1992 09:49:00 PDT
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: Albyn Jones <jones@REED.EDU>
- Subject: Re: interaction effects in regression models
- X-To: A716HOX%HASARA11.bitnet@VM1.MCGILL.CA, STAT-L@MCGILL1.BITNET
- X-cc: A716HOX@HASARA11.BITNET
- Lines: 14
-
- the reviewer is full of hooie (a technical term i learned growing up
- in an agricultural region). "not significantly different from 0"
- is not the same thing as "surely equal to 0"; if you fit a model
- leaving out terms which should be there (eg. main effects which
- are not statistically significant, but still not zero) then the
- estimates of the remaining coefficients that aren't orthogonal to
- those factors are biased, and it may be hard to tell how much the
- bias is. a related issue: if your design involves a blocking factor,
- should you include the main effect for the blocking factor in the
- model even if its non-significant? i'd say yes, because you get an
- inflated estimate of the error variance if you omit the block
- effects from the model: the model should reflect the design!
-
- albyn jones
-