home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!spool.mu.edu!darwin.sura.net!paladin.american.edu!auvm!GRAY.WHOI.EDU!G40_JRG1
- Message-ID: <9211121353.AA10311@aqua.whoi.edu>
- Newsgroups: bit.listserv.stat-l
- Date: Thu, 12 Nov 1992 08:53:44 -0500
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: g40_jrg1@GRAY.WHOI.EDU
- Subject: ANOVA ON FACTOR SCORES
- Lines: 29
-
- From: GRAY::G40_JRG1 12-NOV-1992 08:55:07.87
- To: AQUA::"STA-L@VM1.MCGILL.CA"
- CC: G40_JRG1
- Subj: ANOVA ON FACTOR SCORES
-
- I must comment on the recent discussions.
-
- MANOVA would be appropriate, even on orthogonal rotated factors. Consider
- that the pooled correlation between two factors could be 0 (orthogonal) while
- the correlation in each of two categories were significant of opposite sign.
- The categories did not enter into the extraction of the factors.
-
- If the number of variables, v, is much greater than the number of factors, p,
- then MANOVA on the factors rather than on the original variables has a
- an advantage. The discriminant function used to maximize the differences across
- categories will be sharper. That is the confidence on the parameters of the
- discriminant function will be narrower, and the resultant f-test will have a
- greater power. This comes at a cost: some of the variance of the original
- variables will have been discarded in the v-p factors that were not kept.
- The MANOVA test of differences across categories really tests the hypothesis
- that the v variables were sampled from a single p-variate normal population
- with a given correlation structure (orthogonal or what have you). I am
- thinking purely in terms of principal component factor extraction. I don't
- know if these comments hold for other factor extractions.
-
- Julien R Goulet Jr
- JGOULET@GRAY.WHOI.EDU
- National Marine Fisheries Service
- Narragansett RI
-