home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!stanford.edu!bcm!convex!darwin.sura.net!paladin.american.edu!auvm!MIZZOU1.BITNET!C459006
- Message-ID: <EDSTAT-L%92111122064097@NCSUVM.CC.NCSU.EDU>
- Newsgroups: bit.listserv.edstat-l
- Date: Wed, 11 Nov 1992 17:58:18 CST
- Reply-To: Larry Ries <C459006@MIZZOU1.BITNET>
- Sender: "Statistics Education Discussion" <EDSTAT-L@NCSUVM.BITNET>
- From: Larry Ries <C459006@MIZZOU1.BITNET>
- Subject: robustness and correlation
- Lines: 32
-
- I'm a little confused about the discussion of robustness concerning
- correlation coefficient inferences. As I understand robustness, it refers
- to, essentially, how well a test does when the assumptions about the
- population are violated.
-
- For example, to do a one sample t-test we must assume the data comes from
- a normal population. As I understand it, the t-test works pretty well
- even if the data isn't from a normal population, so we consider it a
- fairly robust test.
-
- On the other hand, if we are even talking about a correlation, we
- basically MUST ASSUME we are talking about a bivariate normal distribution.
- (How many distributions have rho as a parameter? The bivariate normal does,
- and I can't think of any others off hand.)
-
- So my training suggests that you shouldn't look at a corrlation unless
- you have a bi-variate normal. Thus robustness can't be an issue.
-
- On the other hand, we do nonparametric correlations all the time.
- ranks aren't bivariate normal. Hmmmm.....
-
- My point is that basically a correlation has no meaning without the
- assumption of bivariate normality, so maybe people look at correlation
- a whole heck of a lot more than they should.
-
- Any discussion?
-
-
- Larry Ries
- Dept. of Statistics
- University of Missouri
- Columbia
-