Re: Anova

On 23 Jan 2006 08:12:26 -0800, "Thom" <t.s.baguley@xxxxxxxxxxx> wrote:

>
> OK. SNK and Tukey use a pooled error term and therefore aren't ideal
> for repeated/correlated measures.
>
> One approach is to use Bonferroni or modified Bonferroni procedures to
> correct for multiple comparisons. Modified Bonferroni approaches are
> more powerful e.g., the Holm procedure or the Hochberg procedure are
> both easy to follow by hand (the latter is the more powerful). Run the

The "Hochberg procedure" can be *much* more powerful,
because it does not conserve "alpha" but conserves the
"total errors."

It bothers me when people seem to lump Hochberg together
with Bonferroni, Holm, or SNK or Tukey, because that does not
do justice to the distinction. On the one hand, Hochberg offers
much more gain than you could expect between slight variants
in procedure; I think it might be investigated more if that was
always made plain. (I know that I ignored the first couple of
mentions I read because I assumed it would be trivial.)

On the other hand, you should not drop Hochberg into a paper
without noting that it has a novel basis.

> pairwise tests as paired t tests to get the non-pooled error terms.
>
> A better approach might be to use focused a priori contrasts (e.g., for
> trends). These are also raelly just paired tests of weighted
> combinations of means. For example a linear trend for 5 levels uses
> the weights -2, -1,0,1, 2.
>
> The main weakness of the focused contrasts approach is that is doesn't
> work so well for post hoc contrasts. In those cases one can use Scheffe
> ... which is very conservative. Another approach I've seen is to use
> information criteria such AIC to select the most inforamtive pattern of
> differences between a set of means - though doing this properly for
> repeated measures is tricky with current software (but is more powerful
> for post hoc testing than most other aproaches).

--
Rich Ulrich, wpilib@xxxxxxxx
http://www.pitt.edu/~wpilib/index.html
.