Date: Tue, 4 Dec 2001 11:38:57 -0600
Reply-To: "Kevin F. Spratt" <kevin-spratt@UIOWA.EDU>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: "Kevin F. Spratt" <kevin-spratt@UIOWA.EDU>
Subject: Re: Dummy variable adjustment for repeated measures
Content-Type: text/plain; charset="us-ascii"; format=flowed
At 05:56 PM 12/4/2001 +0100, Dave Sorensen wrote:
>I am analyzing a 7-wave, repeated measures (panel study). I have lots of
>independent variables, only one of which I now describe.
>In each of the 7 waves, subjects are asked:
>a) Do you have any close friends?,
>b) if YES, please answer 10 likert-type items concerning the quality of your
>relational bonds with those close friends.
>I sum those 10 likert-type items to create a continuous measure called
>The problem is, 15% of subjects in wave 1, and 10% in wave 2, report having
>no close friends. These subjects therefore have missing values on the BONDS
>summary score in those waves, which means Proc Genmmod ignores all data from
>these subjects in those waves - which is a shame.
>Since being friendless is the mechanism that drives the missing values on
>BONDS, I figure I can retain the rest of the data for these problem cases by
>creating a dummy variable called NOFRIEND which=1 if they have no friends,
>and using it in conjunction with the BONDS scale.
>My question regards what value to give these friendless folks on the BONDS
>scales. I have read it should be some constant, preferably the mean score
>for BONDS among the non-missing cases. But that suggestion was made within
>a cross-sectional context. My context is repeated measures. Should the
>mean constant used be a) the 7-wave mean on BONDS or b) the wave-specific
>mean on BONDS? In the latter case (b), friendless subjects would be
>assigned one of 7 different "constants" depending on wave.
>Rutgers University and University of Copenhagen
If the bonds scale is thought of as some magnitude of the strength of
friendship, then why not have those without close friends at the bottom
of this scale? By putting this group at the mean of the scale for those
with friends, you are suggesting that those without close friends are
average on the bond scale compared to those who report close friends.
That seems idiotic.
The consequence of placing those without close friends at the bottom of
the scale is that you will create a local and possibly a global mode
in the data at the bottom of the scale. This may be handled with a
Kevin F. Spratt, Ph.D. | |
Iowa Testing Programs (ITP) ( )
University of Iowa Back Care (UIBC) \ \_ University
Iowa Spine Research Center (ISRC) \ * \ of Iowa
224-D Lindquist Center \ / Hawkeyes
The University of Iowa |____________ )
Iowa City, Iowa 52242 V
(319) 335-5572 (voice)
(319) 335-6399 (fax)
(319 530-8254 (cell & Voice mail)