Re: ANOVA???
- From: Richard Ulrich <Rich.Ulrich@xxxxxxxxxxx>
- Date: Mon, 21 Aug 2006 18:30:17 -0400
On 20 Aug 2006 13:34:28 -0700, "Jem" <jomilton@xxxxxxxxxxx> wrote:
Hi
I am unsure what statisitical test to do to compare results between
subject and intervention groups in a recent study and wondered if
anyone can help.
I have a group of people with diabetes and a group without, half of
each group is randomised to a 6 week intervention (A) and half not (B),
thus I have 4 groups of completely different people in total.
I hope that this question is hypothetical. This is probably
why other people have not answered by now.
If you are going to recruit subjects for a "trial", you
should have some notion of what you might learn, and
how you might learn it.
Why are there diabetics for one group? Is the intervention
something they need? Are the groups *not* matched on
the important variables?
If groups are not matched at Pre, it is always more difficult to
reach any sort of conclusion, and it is sometimes impossible
to say much useful at all.
As the study is a parallel rather than crossover design I have decided
to look at changes from baseline and compare these between A and B.
However, I am not sure whether to use an ANOVA or non parametric
If the groups differ by much at Pre, then you will have
totally wasted the Control group, if you can't look at the
numbers as similar intervals on a scale.
Are you scales so bad that you think you need to use
rank-transformation?
equivalent to compare the change from baseline between all 4 groups
(diabetics A & B and non diabetics A and B) or whether to do two sets
of t-tests or equivalent so diabetics A v B and non diabetics A v B.
The only thing making me hesitate in doing t-tests is that although the
4 groups are unrelated I expect the diabetics to be more responsive to
the intervention than the non diabetics. Therefore I think doing an
ANOVA with posthoc test would allow be to see whether the change from
baseline in the diabetic A group differs from non diabetic A group
Groups can be compared by
(1) raw outcome = final score);
(2) raw change score = (Post minus Pre); or
(3) "regressed change score" = (Post minus 'expected score').
(1) and (2) are easy to *do*. You can be left with tough arguments
to make them make sense in a report, if the Pre-means are not equal.
Look at the plots of means to see what is going on.
I wrote (3) to include a number of possibilities by saying 'expected'.
Can you expect simple regression toward the overall mean for
both groups combined, or is the situation totally different?
Ancova is okay if Pre means are equal. Also, ANCOVA *might*
still be meaningful and interpretable when the means differ. However,
that requires careful consideration of 'how different' -- in what
direction; where the means fall; what the scale is measuring;
what artifacts exist in measurement... and so on.
If you take Patients at the time of a crisis or episode, and measure
some characteristic extreme during an episode, then the proper
"expected" value would be much closer to their own normal-mean,
*not* to the mean of the sample as it was measured at Pre.
-- And that mean would not be knowable from that present design.
Hope this helps.
--
Rich Ulrich, wpilib@xxxxxxxx
http://www.pitt.edu/~wpilib/index.html
.
- References:
- ANOVA???
- From: Jem
- ANOVA???
- Prev by Date: Re: partner for statistics
- Next by Date: Re: One-sided p-values in regression
- Previous by thread: ANOVA???
- Next by thread: Log-Likelihood with Eviews
- Index(es):
Relevant Pages
|
|
