Date:         Mon, 24 Sep 2001 10:22:37 -0700
Reply-To:     Cassell.David@EPAMAIL.EPA.GOV
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         "David L. Cassell" <Cassell.David@EPAMAIL.EPA.GOV>
Subject:      Re: Repeated Measures with missing data
Content-type: text/plain; charset=us-ascii

Doc Muhlbaier replied:
> I'd use PROC MI and the complete data model (I personally prefer GLM and
> multiple dependent variables, using CONTRASTs to examine the trends.).
Here
> you are assuming multivariate normality.

I strongly agree here.  PROC MI and PROC MIANALYZE give you a way of
dealing with missing values [in some cases, depending on the structure of
the missing values].

> Otherwise, about your only choice is GLM with single dependent variables
and
> the REPEATED statement.  This adds the assumption of homoskedasticity
> (spelling?), one that is rarely met.

I would also recommend evaluating PROC MIXED for repeated measures
problems.
And, if you really have patterns where missing values occur at a point I
for
observation X{I} and then occur at all points after point I as well, then
you
may need to determine whether the Cox proportional-hazards model might be
appropriate for your work.  If so, check out PROC PHREG .

> In either case, you need to look at why the data are missing (MAR, MCAR,
> non-random missings, etc.).  Joe Schafer and Paul Allison at Penn have
> written a lot here that is quite readable.

Lawrence's best point.  The structure of your missing values must be
understood
before you can decide what analyses are appropriate.  In some cases, you
can
easily impute missing values.  In some cases, even multiple imputation will
not
solve your problem.  Only you [and your researchers] know how the missing
values
come about.

David
--
David Cassell, CSC
Cassell.David@epa.gov
Senior computing specialist
mathematical statistician