Date:         Mon, 4 Nov 2002 11:06:46 -0800
Reply-To:     Joël Rivest <jrivest@CDPQINC.QC.CA>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Joël Rivest <jrivest@CDPQINC.QC.CA>
Organization: http://groups.google.com/
Subject:      Resolving a random coefficients model
Content-Type: text/plain; charset=ISO-8859-1

Hi,

I would like to model the growth curve of about 300 pigs using a
random
coefficents model. I plan to include fixed effects like sex and breed,
and some
random effects such as pen and litter. The reason why I want to use
the random
coefficients approach is that animals do not have the same number of
measurements and do not end test at the same weight neither at the
same age.
Using a special covariance structure to model the correlations between
repeated
mesurements on the same animal would probably lead to an artificial
inflexion (when compared to individual growth curve). Thus I would
like to try and use a
random coefficients approach with proc mixed (I could eventually use
proc
Nlmixed and use a gompertz growth curve instead of a polynomial).

Convergence with the random coefficient model is hard to achieve with
a
relativly simple model (1 degre). Adding terms to the model (e.g.
having
polynomial of 2 or 3 degree with the interaction with fixed effects)
would also
be very difficult to solve.

I am wondering if there is a way to simplify the resolution of the
model. Would
there be a way to fit a growth curve to each animal (using a by
statement,
considering only the age as fixed effect), and then analysing the
resulting
individual parameters (considering the errors) to see if there are
fixed
effects on these parameters (random effects as pen and litter would
have to be
considered too)? Hopefully, this would allowed a much more rapid
resolution.


Thanks