home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!pipex!bnr.co.uk!uknet!prd!steve
- From: steve@prd.co.uk (Steve Blinkhorn)
- Newsgroups: bit.listserv.stat-l
- Subject: uses of factor analysis
- Keywords: factor analysis
- Message-ID: <Bz51CI.CKH@prd.co.uk>
- Date: 12 Dec 92 08:38:42 GMT
- Organization: Psychometric Research & Development Ltd
- Lines: 42
-
- The trouble with factor analysis is not that it *never* works, but
- that by and large you can't tell whether it has revealed a deep truth
- about the universe or just produced some nice numbers. There are
- some notorious cases (colour perception is one) where factor analysis
- can produce results that look convincing in terms of the criteria for
- acceptability normally used, but which map not at all onto physical
- reality in any useful way. So while *sometimes* you get a stunning
- vindication, factor analysis (and PCA) doesn't tell you in advance
- that it's done a good job.
-
- Much of the junk in the psychological literature on personality
- measurement (and a lot of it is junk) arises from the deep faith
- earnest authors have in the capacity of SPSS and the like to create
- theory out of data by means of an eigenstructure. And a lot of the
- packaged software around will cheerfully analyse hopelessly defective
- data (e.g. Heywood cases, ipsative data) and produce neat obliquely
- rotated BS.
-
- So get your graduate students to write their own programs (or to use
- SC to build a program out of its elements). That way at least
- they'll develop a conscience. And then start worrying about why we
- think it's sensible to factor analyse correlations *at all*.
- "You mean you collect all this data?"
- "Yes"
- "And then you standardise all the variances to be equal?"
- "Yes"
- "And then you decide which are the important factors on the basis of how
- many standardised variables' worth of variance they explain?"
- "Yes"
- "But you don't have a method of deciding which varaibles to include in
- the first place?"
- "No"
- "And then you publish a paper with just the rotated factor loadings?"
- "Yes"
- "But you publish a list of percentages of variance explained by factor
- number, where the factor numbers are unrotated factors?"
- "But in SPSS, what it gives you is...."
- "What did they teach you in graduate school?"
- "Not a lot........"
-
- ---
- Steve Blinkhorn (steve@prd.co.uk)
-