Rich Ulrich <[EMAIL PROTECTED]> wrote in message news:<[EMAIL PROTECTED]>... > On 8 Dec 2003 11:09:20 -0800, [EMAIL PROTECTED] (Frank Rigous) > wrote: > > > > I am new to the group and not very experienced in statistics. I > > thought I had a classical design, two groups with either treatment A > > or B, measuring the response at 4 time point where the first one is a > > baseline measurement. I was very astonished not finding any example > > for this design. > > Randomized, I hope. Yes, randomized.
> > That sounds classical enough, if you expect a linear trend > and that is what you will test for. Look for "trend". > > > > > My first idea of analysing was to calculate 3 t-tests with the > > difference to baseline. But as I found (e.g. Altmann) this is not > > recommended. He suggested a summary statistic - but this seems not > > possible because the course in time may vary a lot. > > *How* might the course in time vary? At the beginning (between time point 2 and 1) there can be a decrease or increase; afterwards there should be an increase. Therefore at least no linear trend is expected. > > If the advantage is early and sustained, then you get a > powerful test by using the average of the three as 'outcome'. > Usually, the best test of 'final outcome' (the way that you might > describe this one) will be the one-way analysis of covariance. > That compares the regressed-change scores. Most examples I found (thanks!) do not include an interaction. I think it makes sense to include an interaction. Is that right? > > If the advantage is slow and late, you might get the effective > test by ignoring time2 and time3 -- just test time4 with base- > covariate, instead of taking an average of 2, 3, and 4. > > I set up a design for some drug studies where the initial impact > was *strong* and was followed by several periods of > (a) no-more-change, (b) further drift in the same direction, or > (c) drift back towards baseline. That would have been tougher > to test against placebo, but those were big changes and we > were satisfied to describe them: as size of change, and the > subsequent linear trend. I can not imagine how to model this. May you show me the model? Frank Rigous . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
