Date: Mon, 25 Mar 2002 14:09:25 -0800
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: binary and CALIS
Content-type: text/plain; charset=us-ascii
[Manon replied to me directly, rather than to the list. Her message
is below my sig.]
I would recommend that you try this both in PROC CALIS [using Browne's
ADF estimation method, given your large sample size] and using PROC
FREQ. You have a nice set of categorical variables, and your questions
appear to be those which could be addressed using the Cochran-Mantel-
Haenszel general association statistic [use the CMH option to get it
automatically]. If there is a major discrepancy, consider that the
non-normality of your data is likely to be the culprit, and the PROC
CALIS is more likely to be the guilty party.
You also wrote:
> This cohort is representative of the population, and so
> I have to apply sampling weights.
That is [strictly speaking] not true. It depends solely on how you are
evaluating your analysis. You can treat this as a super-population
problem [to use the jargon of Scott Overton] and evaluate the data
without
using the weights. If you actually want to use the weights, then you
should treat this as a probability sampling problem, and also
incorporate
the joint inclusion weights into your analysis. Studies such as those
reported in Sarndahl indicate that not using the full structure of the
sampling design can seriously bias your p-values. Furthermore, properly
incorporating all the design features will require more than a simple
PROC CALIS or PROC FREQ can manage.
Sorry to be the bearer of bad news,
David
--
David Cassell, CSC
Cassell.David@epa.gov
Senior computing specialist
mathematical statistician
-----------------------------------------------------------------------
---------------------- Manon's message --------------------------------
-----------------------------------------------------------------------
Thansk for the answer.
I would like to model something like this:
Breast feeding ------------> health at 5 months ------> health at 18
months --------------> health at 30 months
where breastfeeding is 1 for child was breastfed and 0 not breastfed
health are all 1 for child have poor health and 0 for
excellent health
The above relation is affected by day care center frequentation (which
is
different for 5, 18 & 30 months) where 1 stands for the child went to a
day
care center and 0 for no.
So the final model would be:
Breast feeding
------------> health at 5 months ------>
health
at 18 months --------------> health at 30 months
Day Care at 5 months (line to health at 5 months, 18 and 30)
Day Care at 18 months (line to health at 18 months and 30)
Day Care at 30 months (line to health at 30 months)
Briefly, we followed 2000 children during 30 months at 3 specific time
points (5,18, 30 months of age) to see the relation between
breastfeeding &
subsequent health. This cohort is representative of the population, and
so
I have to apply sampling weights.
We would like to answer to question like "Are breastfed children more
likely to be in better health than non-breastfed children ?", "Or is the
health more influenced by day care center frequentation ?".
We made analysis for each time points (transversal) and saw that
breastfeeding affects health for 5, 18 & 30 months and day care affects
health at 18 & 30 months.
>So I recommend that you *not* use PROC CALIS for your model, unless you
>have a *huge* sample size. [If that is the case, then you should be
>able
>to get away with Browne's ADF estimation method.]
Is 2000 individual is huge ?
>More to the point, why, if all your variables are binary, aren't you
>considering some manner of categorical analysis? You might try
starting
>with PROC FREQ and some likelihood ratio tests. [I can't give much
more
>precise advice, since I have no idea what your data are like.]
Is it better to use a logistic model with repeated measures (how to add
constraints because nothing affect prior a time point) ?
You made good comments yet!
Needless to say that they are very appreciated .
Manon
