Date: Thu, 9 Jan 2003 15:45:04 -0800
Reply-To: Dale McLerran <stringplayer_2@YAHOO.COM>
Sender: "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From: Dale McLerran <stringplayer_2@YAHOO.COM>
Subject: Re: PROC MIXED PROBLEMS
In-Reply-To: <avks5q$7vl$1@woodrow.ucdavis.edu>
Content-Type: text/plain; charset=us-ascii
--- Craig Hadley <cahadley@UCDAVIS.EDU> wrote:
> Hello everyone:
> I've come against a sticky issue that is new to me, and am not
> entirely sure
> how to deal with it. I am trying to test whether 1) children measured
> across
> two seasons show a decline in weight-for-age zscores (WAZ) scores,
Let's pause here for a moment to consider your response variable
which is a z-score measure of weight given age. If you take a
random sample of the population of interest and compute the
z-scores for that random sample and you follow up this random
sample at a later point in time and compute z-scores for everyone,
then the mean z-scores would be the same at time 1 and time 2
since the z-scores must sum to zero. If you have dropouts from
time 1 to time 2 so that only a random sample of subjects is
available at time 2 who were in the sample at time 1, you would
still have, at each time, z-scores which must sum to zero. So,
if you are using random sampling, there should not be any change
in your response mean, right?
Now, if you selected only a subset of the population who at time
1 were high in weight-for-age z-score and then examined their
weight-for-age z-scores at time 2, you would have set yourself
up for problems with regression to the mean.
I don't see how you expect your response to show a non-spurious
mean change over time. Can you clarify this? The only way that
I could see the z-score values declining over time is if you
combine all subjects who are age 14 months at both baseline and
follow-up when computing the z-scores. Then you would have the
situation that for each age group, the z-scores would always
sum to zero. In that case, AGEMONTHS effect in your model
below must be zero. At every monthly age, the mean z-score
is zero.
> and (2)
> whether this decline is greater among children in one group.
This should be testable. The fact that the z-scores should sum
to zero at each time does not mean that the z-scores for each
group should sum to zero. In fact, if you have a treatment
effect, what should happen to the z-scores is that the values
in your treated group decrease while the z-scores should
increase in the untreated group. The z-scores should go in
opposite directions.
> Because
> many
> children were only measured once I'm using proc mixed.
No. Because some children are measured multiple times so that
the errors between some observations are correlated, you are
using proc mixed. You are also using proc mixed because some
observations have other sources of error correlation in that
they are from the same family.
> The issue is
> how to
> appropriately specify the model to accommodate that multiple children
> were
> measured from a household, and children were often measured twice.
> Specifying the random or repeated statement is the troublesome part
> (see
> below). Any help would be great!
> The three things I want to learn from this model are:
> 1. Do children's WAZ scores decline between the two periods of
> measurment
> (SEASON)?
> 2. The magnitude of any decline (size of SEASON coefficient)
> 3. Does this affect children from each group (TREAT) equally?
> (Magnitude of
> TREAT term).
See above. I don't think that 1 and 2 are testable.
>
> I've been using the following code:
> /*
> Code1= Child's unique id code
> Hhid1= Household unique id code
> Treat= Treatment (2 groups) */
> PROC MIXED data=SeasonData covtest;
> Class TREAT sex season code1 hhid1 ;
> Model WAZ= season TREAT sex Agemonths/outp=MyOut;
> /* Specifying the random or repeated statement is the troublesome
> part */
> *Repeated /type=un sub=code1(hhid1); /* Option #1* - I think this is
> correct
> /
> *Random code1; /*Option #2 - pretty sure is incorrect, no household
> term*/
> *random code1 hhid1; /Option #3 * ?/
> RUN;
Before commenting on the appropriate code, please address the
questions about how you will interpret the regression employing
z-scores as the response. I will just state that none of your
repeated or random statements fully addresses all of the
correlation structures which are present in your data. But
before stating other ways to model the correlation structure,
please reply concerning how you are constructing your z-scores
and how you will interpret your model.
Dale
=====
---------------------------------------
Dale McLerran
Fred Hutchinson Cancer Research Center
mailto: dmclerra@fhcrc.org
Ph: (206) 667-2926
Fax: (206) 667-5977
---------------------------------------
__________________________________________________
Do you Yahoo!?
Yahoo! Mail Plus - Powerful. Affordable. Sign up now.
http://mailplus.yahoo.com
