Date: Mon, 12 Jan 2009 09:20:08 -0500
Reply-To: Peter Flom <email@example.com>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Peter Flom <peterflomconsulting@MINDSPRING.COM>
Subject: Re: Repeated Measures, or what?
Content-Type: text/plain; charset=UTF-8
John Whittington <John.W@MEDISCIENCE.CO.UK> wrote
>Following the very helpful responses to my query back in October (my
>original post and Dale's first response below) I now have a couple of
>1...I now have some similar data, but with more measurement time
>points. Instead of just 'Day 0' (baseline), 'Day 1' and 'Day 12', I have a
>baseline ('Visit 0') followed by 'Visit 1' to 'Visit 8' inclusive, the
>visits not necessarily being equally spaced in real time. In view of the
>possible temporary and/or fluctuant nature of responses, there is again a
>desire to undertake hypothesis tests at each of Visits 1 to 8 to compare
>the active and placebo treatment groups with regard to 'changes from Visit
>0' at each of those timepoints.
>Would be appropriate to extend Dale's previous approach to this new
>situation? (does anything have to be done about the multiplicity of
>hypothesis tests?) If so, I confess that I usually get tied up in knots
>constructing the ESTIMATE statements and would very much appreciate
>assistance, perhaps with some general advice on the construction of such
>2...Would the advice be different if the visits were known to be equally
>spaced in time?
>3...As a general interest question ... in pondering the approach which Dale
>suggested before (as below), I wondered whether (in the absence of any
>suggestion that absolute baseline data might influence responses) anything
>would be lost by applying the analysis to 'Changes from baseline' data,
>rather than to the raw data?
I will let Dale speak to any modifications of his particular solution. My impression is that his general approach is still good, but there may be many more tests.
An alternative, or additional, idea would be to forget about the contrasts between specific time points, and estimate slopes over time for the different groups, controlling for such variables as you wish. This would yield considerably less output to interpret. You might want to look at quadratic trends over time, or even cubic. I'd *strongly* suggest first doing some plots (were it me, I would do these in R; people who are expert in SAS Graph could show you how to do them in SAS Graph) .... e.g. look at each person's results over time. If N is very large, you might look at a random sample of people, or split the results over several graphs.
But the general approach of using slopes seems to me (and Dale, or others, feel free to correct me!) not only to be closer to the usual use of PROC MIXED, but to answer a more relevant question. I think your interest is in whether one group is getting better faster than the other, if so, how much faster, and how the rate of improvement can be characterized.
PROC MIXED deals very well with unequal time points.
Changes from baseline is a different question - I think the key point there is which question you are interested in, substantively.
Peter L. Flom, PhD
www DOT peterflom DOT com