Date:         Thu, 2 Nov 2006 00:01:46 -0500
Reply-To:     Rich.Ulrich@comcast.net
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Richard Ulrich <Rich.Ulrich@COMCAST.NET>
Subject:      Re: Missing covariate at some group levels
Comments: To: sas-l@uga.edu
Content-Type: text/plain; charset=us-ascii

On 1 Nov 2006 09:29:41 -0800, hoenwork@yahoo.com wrote:

> Hi, I have an ANCOVA with independent variables TREATMENT, START_TIME
> (1,2), BASELINE_WEIGHT1 (missing for start_time group 2), and
> BASELINE_WEIGHT2 (missing for start_time group 1).

I think I assume that Final_weight is the dependent variable, but
that is not critical.

>
> Does anyone know how to model it, if it can be done?  One idea is to
> set the missing values to the mean of the non-missing values of each
> variable.

The conventional way to put in a variable that has Missing is
to use <anything>  for the missing value and to include another
dummy variable, which flags that the variable is missing.  Using
the mean will minimize the 'effect' of the flag, but you might
look at it both ways to make sure.

In this case, you have two variables which (it looks like) can
be signaled by the same Dummy, namely, Start_Time.


>         Then the baseline means that treatment means adjust to is
> correct, but the regression slope between the covariates and the
> outcome is off a little, because of including the contrived points
> (i.e. because of substituting in the means where there are missing
> values).  The contrived points would be of low influence, but still
> bias the answer.

--
Rich Ulrich, wpilib@pitt.edu
http://www.pitt.edu/~wpilib/index.html