I'm going to go quickly over this example--I just want to be sure
that people understand the basic issues and the design of the
analysis.
Carter, J. C., & Fairburn, C. G. (1998).
Cognitive-behavioral self-help for binge eating
disorder: A controlled effectiveness study. Journal
of Consulting and Clinical Psychology, 66,
616-623. The data are available at carter.sav. (The individual
data values are not integers because of the way that
I generated them, but they have the appropriate means
and variances.)
The authors wanted to look at some of the
self-help opportunities, on the grounds that there
will never be enough therapists to treat everyone who
needs treatment. They wanted to compare the
"pure" self-help condition with a
"guided" self-help condition. They had a
Waiting-list group, but I won't use them here because
they were measured at only two times.
They recruited 69 subjects who were divided into a
guided self-help group (n = 34) and a pure
self-help group (n = 35). They recorded data
from these groups at pretest, posttest (12 weeks)
3-month follow-up, and 6-month follow-up. (It is not
clear whether these are the correct sample sizes or
whether there should be about 12 more per group, but
I think the numbers are correct. This is important
for power.)
The pure self-help group was mailed a copy of a
book (Overcoming Binge Eating) by one of the authors
and told to read it. That's all.
In the guided self-help group the subjects came in
for 6-8 25 min. sessions "in which the
facilitator supported the participant in the use of Overcoming
Binge Eating." (That's all they said,
except to go on at some length about how little
background the facilitators had in therapy.) (Hard to believe this
study was published in that journal.)
The dep. var. that I will use is the number of
binge eating episodes/28 days. The means and standard
deviations follow. (These are their values. I set the
correlations between .40 and .60.)
Note the obvious trend for subjects to show
reduction over time.
Note the slight tendency for the Guided group to
do a bit better than the Pure group.
No apparent interaction.
Plotting the data
The pretest and posttest scores for the waiting list groups
had means of 21.6 and 13.5.
Note that the wait-list control, which is not
part of our analysis, also showed a substantial
drop over from pre-test to post-test.
What does
that cause us to worry about?
What generality might that kind of effect have?
Subjects were about 40 years old, and
reported an average onset of binge eating of
23.6 years. (Mean weight was 85.8 kg.)
Covariance Matrices
(Note that these are labeled as if they
were correlations, but they are covariances. This is simply the
result of how I used the pivot table options.)
We can see that we have some problems, both with
the covariances and with the variances, especially
with the pure self-help group.
We will deal with this shortly.
Overall analysis
I have changed the order of the printout for
discussion purposes.
The Between-group difference is not significant,
although it was for them.
This gives me a good way of talking about
Bradley and Orfly's results, which show that the
power of the between-subject
effect is reduced when there is a
correlation between the within-subject measures.
(Really Bradley and Russell, 1998, Behav.
Res. Meth., Inst. and Comp.) Elaborate on
this. (We usually think of repeated measures
designs as more powerful, when in some cases they
are not.)
Perhaps I set my correlation too high, because
my means and standard deviations match theirs
exactly, but I don't get that between subject
effect.
What would we conclude if we had a Group effect? How would
that influence our overall interpretation of the study?
Note the significant effect for Time, but the lack
of an effect for the interaction.
Comment of correction factors for sphericity.
Trend
The within-subject analysis is always followed (by
default) with a trend analysis. This is shown below.
Explain what this means.
There is a cubic effect for Time, but I
certainly don't see this in the data.
(A linear interaction effect would mean that there
is a different linear effect for the two groups.
Same with quadratic. What if one group is linear
and the other is quadratic?)
Manova
The standard Manova printout follows.
Review what I said the other day about the nature of Manova.
Comment on why we might want to use Manova here. Comment on why we
might not.
Explain what a Manova represents.
The different tests are all the same here because
we only have two groups. That would not be the case
if we had more than two groups.
Point them to the chapter to read about Manova.
Simple Effects
The simple effects are pretty obvious here, but
they aren't always.
Between-subjects simple effect
To be safe, we would want to make sure that
the groups did not differ at the start of the
study. This means we would compare GSH and PSH at
time 1.
This is just a straight one-way Anova on
Binge1.
As we would hope, there are no differences
between the groups..
If we had had a significant group effect, we
might want to run a similar test at post-test and
at the follow-up times. The test would be done
exactly as we have done this one.
Within-subject simple effects
I think it would be silly to run any for these
data, but I'll do one just for show. I'll only
use the GSH group
I split the data using the Data/Split
file command. I also asked for simple
contrasts, using the baseline time as the
reference.
Simple contrasts contrast each mean with a reference mean,
which can either be the first or the last. I used the first.
The contrasts
It is apparent that every time is different
from baseline.
Discuss the issue of error terms.
By splitting the file, the data from the
PSH group does not enter into this analysis,
which saves us worrying about G-G and H-F
(except so far as their corrections for
these data are needed). In other words,
any problems we have that came from the PSH
group are not messing things up here.
For the simple contrasts, you can see that
individual error terms are computed. Again,
we are not in danger of running into problems
with sphericity. This is the best way to do
these analyses.
