On Sun, 06 Oct 2002 18:00:47 GMT, Paul Victor Birke <[EMAIL PROTECTED]> wrote:
> Herman Rubin wrote: > > In article <002701c26c2e$e748f990$55e8b23f@DHJC4K01>, > > Karl L. Wuensch <[EMAIL PROTECTED]> wrote: > > > >> I try to stress to my students the importance of screening their > >>data carefully for out-of range values, outliers, etc. prior to any other > >>analysis. I demonstrate to them how a single seriously out-of-range value > >>can ruin an analysis. .. HR > > > > However, what you call an outlier can be the most important > > part of the data. One should not just call something an > > outlier because it is not too close to the main body; it > > is necessary to justify rejecting the datum. ... PVB > > Exactly. In fact, it can be gold. > > Treatment of outliers must be done very carefully. This I should > mention is also stated in Prof. R. Montgomery in his book Design of > Experiements. > > However, that being said, one should make an effort at preprocessing and > removal of obvious outliers. Those out-of-range, impossible, values have to be removed. If a single value (or two) will totally degrade the ANOVA, that is justification, as I see it, to "do something else." If scores are usually in the range of 1 and 2, then a score of 100 will account for too much "variance" - squared deviations - for the little deviations around 1 to carry information. More broadly: "means" are intended for useful comparisons. Explicitly or not, you intend to compare parameters, or (something like) central tendencies. If outliers make nonsense of the intention, then you ought to "do something else" - which always includes, Document Whatever You Do. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
