Date: Mon, 9 Mar 2009 11:40:29 +0100
Reply-To: Thomas Fröjd <thomas.frojd@NEURO.UU.SE>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Thomas Fröjd <thomas.frojd@NEURO.UU.SE>
Subject: Adjusting data when measurements are taken at two different
Content-Type: text/plain; charset=ISO-8859-1; format=flowed
Hi SAS list.
Once again I have a problem I hope someone can give me a hint on.
I am trying to analyze the differences between two groups in the result
of a psychometric survey after they have been affected by a traumatic
event. The survey they have filled in consists of multiple standard
psychometric scales designed to measure stress symptoms. Group A has
taken the survey at 14 months after the event and then retaken the same
survey after 36 months. The mean scores in the later survey is lower but
that is expected since the effect of the event is expected to decrease
with time. I think it is a reasonable assumption that the scores
decrease in a linear way.
The trouble is that group B has taken the survey at 21 months from the
event wich make the two groups not directly comparable. My goal is
therefore to adjust the scores for group A so I can make a regression
model with demographic and other factors as covariates and the scores of
both groups at 21 months as the dependent variable.
My first idea was to simply take the proportional change in scores for
each month between the 14 months and 36 months study and then adjust the
scores observed at 14 month down to what they could be expected to in
21 months. I think this is fine if the scores are truly continious and a
linear slope is assumed. However I think there are some problems with
the first assumption.
First, the scores are not truly continious since they are the sums of
yes/no questions for diffrent symptoms where "yes" is scored 1. Since
the total score for an individual is supposed to measure stress and the
questions are not weighted in any way the total sums can't be treated as
Furthermore the researcher I work with wants to dichotomize the total
scores before analysis according to cutoffs that are standard in her
field. (The whole questionaire and scoring system is a standard
psychometric scale and is supposed to be treated like this). She then
want to use logit analysis to estimate odds ratios. The trouble with
this is that many of the individuals have exactly the score above the
cutoff. If I then use the adjustment i wrote about above (meaning an
adjustment of about 0.8 times the 14 months scores) theese individuals
would be put below the cutoff and the incidence of stress symptoms would
go from something like 40% to about 25%. Just by intuition I think this
would be wrong since this drop seems way to large when the underlying
scores only where supposed to drop 20 %.
I understand that there is impossible to give a clear answer for the but
I would be happy to hear your experience if you encountered similar
Anothed thing I was thinking about was if I could use a mixed model and
set the estimates to show at 21 months. Then I dont have to adjust the
scores at all but maybe this would be wrong since there is only one
measurement. for the B group and no data for 21 months for the A group.
All comments are greatly appreciated.