BICKIS@SASK.USASK.CA (Mikelis Bickis) (02/08/90)
Reply-to: "Evan G. Cooch" <COOCHE@QUCDN.BITNET>
.
.
.
> However, this would seem to me to be appropriate if I was A PRIORI
> deciding to test (say) slope A vs C alone. But in general, when I
> have significant heterogeneity in slopes (significant CLASS*X term),
> I'm not particularly interested in A PRIORI testing any two slopes,
> but A POSTERIORI testing ALL slopes to see which ones are different
> from which other ones. In simple ANOVA, I would use Tukey-Kramer, or
> something analogous.
>
> Given that paired-slopes testing is not typically a standard procedure,
> are A PRIORI adjustments appropriate? If so, how (and what) do you adjust
> in the tests?
>
> My gut feeling is that since I'm testing everything, such adjustments may
> be unecessary.
Sorry Evan, you have it backwards. It is precisely *a posteriori* tests
that require adjustments for multiplicity. So if you are testing *all*
slopes, you *need* an adjustment such as Bonferroni, Sidak, Tukey, etc.
Bonferroni is good because it is simple, always possible, and gives good
sensitivity if the number of tests isn't too great. Scheffe tests are
also always available, but not simple, and may be less sensitive for a
few comparisons. Tukey's tests can be used if all the slopes are based
on the same design-- however, if you are fitting a common intercept, you
have to put in an adjustment for the correlation. (Basically, instead of
giving it the variance of an estimate, you give it one-half of the
variance of the difference.)
Mik Bickis
Dept. of Mathematics
University of Saskatchewan
Saskatoon
BICKIS@SASK.BITNET
bickis@sask.usask.ca
<( I use SAS 5.18 running on a VAX 6300 cluster under VMS 5.1 )>