home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!paladin.american.edu!auvm!VTVM1.BITNET!WLTPIRIE
- Message-ID: <EDSTAT-L%92112016321820@NCSUVM.CC.NCSU.EDU>
- Newsgroups: bit.listserv.edstat-l
- Date: Fri, 20 Nov 1992 15:34:08 EST
- Reply-To: Walt Pirie <WLTPIRIE@VTVM1.BITNET>
- Sender: "Statistics Education Discussion" <EDSTAT-L@NCSUVM.BITNET>
- From: Walt Pirie <WLTPIRIE@VTVM1.BITNET>
- Subject: Diff in r**2
- Lines: 38
-
- The original question was about a test of whether two R**2 values differ.
- Frank Dane suggested adding a dummy variable to account for the two models. As
- suggested by another respondent, that answers a different question: i.e. not wh
- ether the R**2 are different, but whether the models are different.
- IMHO, the latter is a very natural question to ask, and the dummy variable is
- an excellent way to answerit. The one constraint is the assumption of equal var
- iance between the two samples, but that could easily be handled by weighted lea
- st squares. Also IMHO, the original question is a "questionable" one in the fir
- st place. In almost any modern regression book, one can find comments to the ef
- fect that R**2 is not a reliable measure of anything, and is humongously
- overused and misused. Also, from a formal standpoint, as far as I've ever seen,
- the distribution of R**2, even asymptotically, is unknown so that formal infer
- ence using it is not possible. I'd be interested if anyone knows more about tha
- t.
- Finally, another respondent suggested a MANOVA approach which equates correlati
- on and regression. While it is true that mathematically R**2 is the square of t
- he correlation, I've always been taught, and believe, that if formal inference
- such as testing is involved, regression and correlation cannot be casually inte
- rchanged because of the different assumptions underlying the validity of the pr
- ocedures. In otherwords, you can't test R**2 by pretending it's just a (square
- of) a correlation coefficient. And I don't think you'll see any regression
- publication by a statistician supporting that approach.
- If it were true, then the distribution of R**2 would just be the distribution o
- f the square of rho, about which quite a bit is known, and tests of R**2 would
- likely be a common (bad) procedure.
-
- |==================================|============================|
- | Walter R. Pirie | |
- | Department of Statistics | |
- | Virginia Tech | |
- | Blacksburg, VA 24061-0439 | |
- | | |
- | Tel. 703-231-5441 | |
- | Bitnet WLTPIRIE@VTFVM1 | |
- | Telnet WLTPIRIE@VTVM1.CC.VT.EDU | |
- | | |
- | Fax 703-231-3863 | |
- |==================================|============================|
-