One point should always be emphasized in the discussion of this topic, and that is that Scheff�'s procedure is one-to-one with the F-test. If the F-test is significant, then there is some contrast that is significant, and vice versa. The significant contrast may not be a pairwise contrast, which is often a point of puzzlement to many.
If only pairwise contrasts are of interest, then one is well advised to test all interesting ones with Tukey's procedure; and to ignore the F-test, unless the F-test probability is so far from the desired alpha level that pairwise testing will be bootless. An F-test is a test of many uninteresting hypotheses, and thus is considerably less sensitive to pairwise contrasts than is a direct pairwise test: this lack of sensitivity is a general critique of Scheff�'s procedure. Rich Ulrich wrote: > > On 19 Dec 2002 04:55:35 -0800, [EMAIL PROTECTED] (Ivan Balducci) > wrote: > > > Hi members, > > Please, may someone explain to me this doubt: > > why is not right to perform a Tukey HSD test > > after an ANOVA came out non-significant ? > > Why Tukey HSD test should be done only following an ANOVA that was > > significant ? > > Which is the reason ? > > If you are doing LSD (simple t tests) as followup testing, > it is "followup": By definition, you need the overall test if > you are going to protect against excessive alpha error. > I don't know which of the other procedures are actually > 'posterior tests' in this same rigorous sense. > > As Zar says, "Although not actually required by theory, > multiple comparison testing is most commonly performed > only if an analysis of variance first rejects a multisample > hypothesis of equal means." [Biostatistical Analysis, 1999.] > > For most studies, I think it is going to be a pretty-much- > irrelevant matter of caution, since the extra step won't > make any difference. (In the instance of the Sheffe tests, > it *can't* make a difference in the usual direction. Sheffe's > insists that each pair can account for an overall effect. > On the other hand, you might have an overall effect and > *not* see any of your pairs look significant by Sheffe's.) > > For the Tukey HSD test, the reason is mainly precedent, > combined with that caution, or with the statistical illiteracy. > > -- > Rich Ulrich, [EMAIL PROTECTED] > http://www.pitt.edu/~wpilib/index.html -- Bob Wheeler --- (Reply to: [EMAIL PROTECTED]) ECHIP, Inc. --- (302) 239-6620, voice FAX 724 Yorklyn Rd., Hockessin, DE 19707 Randomness comes in bunches . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
