home *** CD-ROM | disk | FTP | other *** search
- Comments: Gated by NETNEWS@AUVM.AMERICAN.EDU
- Path: sparky!uunet!stanford.edu!bcm!convex!darwin.sura.net!spool.mu.edu!yale.edu!news.yale.edu!mars.caps.maine.edu!maine.maine.edu!cunyvm!psuvm!auvm!UNC.BITNET!UPHILG
- Message-ID: <STAT-L%92111711305709@VM1.MCGILL.CA>
- Newsgroups: bit.listserv.stat-l
- Date: Tue, 17 Nov 1992 09:10:00 EST
- Sender: "STATISTICAL CONSULTING" <STAT-L@MCGILL1.BITNET>
- From: "Philip Gallagher,(919)966-1065" <UPHILG@UNC.BITNET>
- Subject: Do assumptions violations mean anything?
- Lines: 64
-
- The recent discussion of assumption-violations has reinforced
- my long standing (not very popular) observation that statisticians
- and others attempting to practice statistics often get so
- comfortable with our (valid) jargon and (again valid) knee-jerk
- practices that we gloss over, perhaps even in our own minds, any
- real meaning that our practices may be accessing. I illustrate
- with a one-way ANOVA. So we do some kind of test for homo-
- scedasticity; it fails. What does that mean? Not "The data
- violate the assumptions", not really. What it means is that we
- have one or more cells in which the distribution of the data differs
- from the distribution of the data in other cells. If that were
- true, what real meaning would an analysis that certifies the cell
- means are not equal have? Suppose the distribution in one cell
- were highly skewed to the right and another cell to the left,
- but with equal means and equal variances? It would almost
- certainly be a very unusual set of scientific data indeed where
- the failure of an ANOVA to detect differences in means would
- be the result the scientist wanted to have called to his attention.
- It strikes me that the testing of assumptions is most easily taught
- by showing the students that violation of the assumptions means
- that the distributions differ, but not necessarily in the way that
- the specific procedure (say, ANOVA) is directed at (equality of
- means). Once the student sees clearly the underlying phenomon
- that leads to an assumption violation it becomes very hard to
- prevent the student from looking for those violations (and in a
- very perceptive way, too).
-
- I have had amazingly good experiences in the last few years by
- encouraging students to "look for systematic characteristics in
- those persons for whom the model does not predict well" rather
- than "examine the residuals". My first real success along this
- line came after having begged, cajoled, and demanded that a
- osteoporosis student examine the residuals in the model for three
- months and gotten nowhere; when I said "Well, forget about looking
- at the residuals, just figure out which groups the model doesn't
- fit well" I got the best part of the answer in (gasp!) two hours.
- Two months later over a celebratory drink I flabbergasted the
- student by explaining that she had actually done an analysis of
- the residuals. (There is success in this world sometimes - this
- person is now faculty at another school, and last month one of her
- students was complaining to me about being forced to examine
- residuals! Hallelujah!)
-
- The gist of this not extensively edited comment is that we often
- become so entranced by the mathematical aspects of what we are
- doing that we fail to remember that the mathematics (at least
- for statistical analyses) is usually a reflection of some aspect
- of the data that one need not be a statistician to understand.
-
- I conclude with my favorite way of examing the similarity of two
- distributions, attributable to Dana Quade. One plots the empirical
- distribution of distn1 on the Y-axis against the empirical distribu-
- tion of distn2 on the X-axis. (When not looking for differences
- in location I center both distributions at zero first). If the
- distributions are similar, the result will be a straight line (at
- 45 degrees if you scale the axes cleverly). Large differences
- in dispersion result in S-shaped pictures; the graph is very
- informative, both to the statistician and to the scientist.
- Sometimes this picture makes the differences in the distributions
- so clear that everyone gladly abandons the original intention of
- testing means. Thank goodness.
-
- Phil Gallagher
- uphilg@unc
-