Date:         Mon, 20 Jun 2011 08:42:17 -0500
Reply-To:     Robin R High <rhigh@UNMC.EDU>
Sender:       "SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
From:         Robin R High <rhigh@UNMC.EDU>
Subject:      Re: DF in mixed models with repeated measures
Comments: To: Nuria Chapinal <nchapinal@YAHOO.COM>
In-Reply-To:  <201106181308.p5IAkjjW010505@waikiki.cc.uga.edu>
Content-Type: text/plain; charset="US-ASCII"

For repeated model covariance structures, the kenward-roger option with
ddfm=kr is often considered a better choice, so I would suggest entering
it. DF computations are rather complex depending on the variances which
depend on the option entered for type= , plus with both RANDOM (default is
type=contain) and REPEATED (default is type=bw) statements further
complicate the issue.  Regarding 2 of your 4 selections, the one with a
REPEATED having cov type=cs along with the RANDOM statement is not
recommended because under positive covariance, they both produce the same
CS matrix.  And the one with type=un is estimating components for an 8x8
matrix which is likely imposing too severe of a structure, esp. with only
272 observations.  I'd also look at results from type=ante(1) and arh(1)
along with the first two and also consider which one produces the
substantially smaller value for BIC.

Robin High
UNMC





From:
Nuria Chapinal <nchapinal@YAHOO.COM>
To:
SAS-L@LISTSERV.UGA.EDU
Date:
06/18/2011 08:10 AM
Subject:
DF in mixed models with repeated measures
Sent by:
"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>



Good morning!

I have a dataset with 34 cows, 10 belonging to the treat mast=1, 24
belonging to the treat mast=0. Each cow has 8 consecutive days of data
(feeding time, in particular, in the example below). So, in total I have
272
obs. I want to consider day as continuous because I am interested in the
slope. Match is a block that I considered random, and there are 10
different
ones. Each block has a cow from the treatment mast=1, and from 1 to 3 from
the treat mast=0. Every cow has a unique ID number, so I didn't nest in
the
repeated statement.

I ran the model 4 times, with different covariance structures. My question
is: why are the residuals DF for day so different when I use the different
structures? It looks like the difference is driven by the fact that in
some
cases, the total DF for the effect of day is 272-1, because each cow has 8
observations, and in some others the total DF for the effect of day is
34-1,
because each cow has 1 single slope? What's the correct option? Should
every
cow contribute to multiple or 1 DF to the total DF of the day and day*mast
effect? It day was class, they would contribute to multiple, but since day
is continuous, I am confused. Why are the different covariance types
dealing
with DF so differently?

I can send the output by email.

The models are (toep yields the smaller AIC):

proc sort data=pilar;
by cow day;

proc mixed data=pilar covtest cl  noitprint  noclprint ;
class cow match mast;
model min =  day mast mast*day/ solution ddfm=satterth outp=p;
repeated /subject=cow type=toep;
random match;
run;

proc mixed data=pilar covtest cl  noitprint  noclprint ;
class cow match mast;
model min =  day mast mast*day/ solution ddfm=satterth outp=p;
repeated /subject=cow type=ar(1);
random match;
run;

proc mixed data=pilar covtest cl  noitprint  noclprint ;
class cow match mast;
model min =  day mast mast*day/ solution ddfm=satterth outp=p;
repeated /subject=cow type=cs;
random match;
run;

proc mixed data=pilar covtest cl  noitprint  noclprint ;
class cow match mast;
model min =  day mast mast*day/ solution ddfm=satterth outp=p;
repeated /subject=cow type=un;
random match;
run;


Thanks!