Date:         Mon, 31 Mar 2003 12:29:39 -0500
Reply-To:     Steve Albert <salbert@AOL.COM>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Steve Albert <salbert@AOL.COM>
Subject:      Re: conceptual statistical question

Jian,

You haven't specified analyses here.

Option 1 gives explanatory variables, but doesn't say what you're using as
the dependent variable -- I would assume it's score for each individual
patient at each time point, but you don't say.  I suspect you're
essentially fitting a mean to each cell of a 2x2 table classifying the
observations by 2 time points and 2 treatment groups.  I'd also suspect
that you're looking at the change pre- to post- in each group, and
comparing the changes to see whether the change in the treated group is
significantly different from the change in the untreated group -- but
again, you really haven't specified what you're trying to test.

Option 2 specifies a dependent variable, but doesn't say what explanatory
variables you include in your model.  I'd guess you are using group as the
only explanatory variable, since time is subsumed into the change measure.

Option 3 adds another explanatory variable to option 2 -- but the original
list of explanatory variables is still missing.

If my assumptions are correct (or at least close), look at the sources of
variation in options 1 and 2.  Option 1 would seem to include both
variation within patients over time and variation between patients; option
2 removes the baseline variability between patients by looking only at the
change within each patient.  Depending on exactly how you're specifying
your analyses, option 2 may be more powerful.

Comparing 2 and 3, the first question I would have is what do the subject
matter experts expect?  If the size of the treatment effect may vary
depending on what the initial level of your dependent variable was, then it
may make very good sense to allow for this in your model, and a reasonable
first pass might well be to include the baseline measure as a covariate.

Hope this helps.

Steve Albert
Director of Biostatistics
Spectrum Pharmaceutical Research Corp.
San Antonio, TX
SAlbert at SpectrumCRO dot com


On Mon, 31 Mar 2003 16:12:13 +0000, Jian Mao <maojianj@HOTMAIL.COM> wrote:

>For pre-post and 2 group (trt vs. control) data, which analytic approach is
>better or more appropriate?
>
>1. group, time, and group by time interaction
>2. change score
>3. change score, but also includes pre-test score as covariate.
>
>I thought 1 and 2 are equavalent.  Some people use 3, but I am not sure
>which approach is more appropriate.
>
>Thanks for your comments.
>
>Jian