to sci.stat.edu and jimkk.
On Fri, 05 Apr 2002 12:08:03 GMT, [EMAIL PROTECTED] wrote:
> Hi, I just did a search on google, tinking this must be a common
> question, but didn't find much.
>
> I'm doing a multi-factor repeated measures ANOVA, A(5) x B(3) x C(2).
>
> I'm getting big main effects as expected, but the interaction is what
> is of most interest, particularly between levels 4 & 5 of factor A,
> and levels 3 & 4 of factor B.
>
> I believe polynomial contrasts involve the use of coefficients in
> order to test for linear (and above) trends. But I can conceive how
> the coefficients could be structured to look at the differences
> between two levels of a 5-level factor. But what I can not imagine is
> how to use it to compare interaction between two factors.
If I understand the problem: you have a w/i subject design,
and you want to look at one particular interaction, that is,
between particular cells, (4 and 5) in factor A, (3 and 4) in
factor B. [ I note in passing that the problem states "B(3)"
and I hope this does not denote a signal failure in my
understanding.]
You certainly don't want to try to extricate the meaning of
a few cells, by reading high-order polynomial trend lines.
If there were no factor C, you could be happy with a paired
t-test. That would contrast (4,3)+(5,4) versus (5,3)+(4,4),
the diagonals of the subtable. You would incorporate
both levels of factor C for each of the items named.
To take factor C into account, the simple solution seems to
be to rerun the analysis with just the 2x2x2 scores that matter.
>
> I'm using SPSS and when doing my fully repeated measures design, it
> does not offer me the option of doing post hoc tests. I think I've
> read that one can not do post hocs on repeated measures, yet messages
> in this newsgroup have indicated otherwise (on google).
The formal problem for so-called post-hocs is (a) the routines
with names were originally derived in terms of pooled variances.
Assuming that turns out to be, usually, overly hazardous
for 1-way repeats ("sphericity") and flat wrong for the multi-way.
The usual followups are paired t-tests, with ad-hoc corrections
for the number of tests.
Hope this helps.
--
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/ .
=================================================================
