home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!stanford.edu!bu.edu!olivea!spool.mu.edu!uwm.edu!psuvax1!psuvm!auvm!UNCVM1.BITNET!WDAVIS
- From: WDAVIS@UNCVM1.BITNET (Walter Davis)
- Newsgroups: bit.listserv.stat-l
- Subject: Re: the meaning of transformations
- Message-ID: <STAT-L%92111718483609@VM1.MCGILL.CA>
- Date: 17 Nov 92 23:30:51 GMT
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- Lines: 45
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- In-Reply-To: Message of Tue,
- 17 Nov 1992 15:55:45 +0000 from <BS_S467@NEPTUNE.KINGSTON.AC.UK>
-
- On Tue, 17 Nov 1992 15:55:45 +0000 Stephen Gourlay said:
- >In the discussion of testing assumptions, someone mentioned transformations.
- >
- >I have always had a problem with these, namely, that whilst I know that my
- >data represented e.g. counts of something, I cannot figure out a
- >non-mathematical understanding of what transformed counts mean. As a result
- >I am very reluctant to transform data just to 'make it fit'.
- >
-
- I tend to agree with Stephen that transforming variables to the point
- of uninterpretability is generally not useful. For example what would
- be the substantive interpretation of the effect of the arccosine of X
- on the eighth root of Y. You may laugh, but I TA'd for a grad stats
- course in which for the final we gave them a mystery data set and
- one student actually tried an eighth root transformation (and,
- naturally, the variables actually came from a normal distribution,
- it was just a somewhat odd-looking sample).
-
- There are some transformations with useful interpretations however.
- For example, log transformations allow for interpretation in terms
- of elasticity (or percentage changes). If the model is of the form
- lnY=BlnX then a 1% change in X leads to an expected B% change in Y
- (lin-log and log-lin have variations of this interpretation). Also
- odds or logit transformations of proportion variables have good
- substantive interpretations and often have 'better' distributions.
-
- I'm not ruling out complex transformations, I'm simply trying to say
- that a simpler transformation (or none) which gives a nearly as good
- fit as a complex one is often preferable especially with regards to
- substantive interpretation. If a complex transformation is needed,
- one means of interpretation is to pick a set of meaningful values
- for the variables (in their raw forms), work through all the
- transformations, then produce a predicted value. Then change one of
- the variables one raw unit, reapply the transformations and see
- how large the shift in the dependent variable is. In general you
- may want to pick more than one set of values at which to evaluate
- the effects. This technique is fairly simple when your main
- interest is in the effects of a small set of variables.
-
- hope this helps,
-
- Walter Davis <WDAVIS@UNCVM1>
- Department of Sociology
- Institute for Research in Social Science
- UNC - Chapel Hill
-