home *** CD-ROM | disk | FTP | other *** search
- Path: sparky!uunet!stanford.edu!lll-winken!elroy.jpl.nasa.gov!sdd.hp.com!swrinde!news.dell.com!paladin.american.edu!auvm!MATH.UNB.CA!ROLF
- From: rolf@MATH.UNB.CA ((Rolf Turner))
- Newsgroups: bit.listserv.stat-l
- Subject: Re: Data snooping.
- Message-ID: <9210142133.AA24073@halmos.math.UNB.CA>
- Date: 14 Oct 92 21:33:21 GMT
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- Lines: 36
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
-
- I have the impression that quite a bit of the recent discussion on
- data snooping has been off-base and mis-represents what the phrase
- ``data snooping'' actually means.
-
- Exploratory data analysis --- looking for patterns and making
- discoveries --- is not data snooping. Data snooping consists in
- formulating and formally testing hypotheses based upon one's
- exploration of the data rather than upon a priori conjectures. As
- such it renders probability statements associated with the results of
- such hypothesis tests essentially meaningless.
-
- Someone in the discussion has suggested reporting the p-values of
- such post facto hypothesis tests as a form of descriptive
- statistics. There is clearly no intrinsic harm in so doing, as long
- as these p-values are clearly distinguished as being ``descriptive''
- rather than having a genuinely probabilistic meaning. However, the
- general devotion to p-values and 0.05 significance levels being what
- it is, it seems to me that such a practice runs a serious risk of
- being misinterpreted. (The warning that these p-values are not
- ***really*** p-values will be treated as ``fine print'' and ignored.)
-
-
- cheers,
-
- Rolf
-
- Rolf Turner
- Dep't. of Maths and Stats
- U.N.B., Fredericton, N.B. Canada E3B 5A3
- rolf@math.unb.ca
-
- ----------------------------------------------------------------------
- | |
- | Nice day, isn't it? |
- | |
- ----------------------------------------------------------------------
-