Date: Tue, 8 Jun 2010 09:48:09 -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: Using the CONTRAST Statement in PROC MIXED for 3 fixed effects
In-Reply-To: <201006081148.o58AkqCD027660@willow.cc.uga.edu>
Content-Type: text/plain; charset="US-ASCII"
Anne,
This question leads into one of my favorite topics, but before I
demonstrate it, will quickly show the "usual" way of writing a contrast or
estimate in MIXED. Your verbal descriptions assume the three-factor
interaction is necessary. I also would further examine the repeated
correlation structure, esp. with both a RANDOM and REPEATED statements
entered as given, but that is beyond the scope of this note.
proc mixed data = t;
class subject treatment group time;
model hydration = treatment group time;
repeated group time / subject = subject type = un@un;
random subject/ subject = subject;
run;
First, note the order of variables placed on the CLASS statement and the
order and number of their levels
Class Level Information
Class Levels Values
treatment 4 10 20 30 K
group 3 1 2 3
time 2 24h 3h
The order of the levels for treatment and time may be different from what
you expect,
but that is how SAS interprets your coding. Here 3h implies level of time
= 2
From top to bottom, write out the levels in the following heirarchy:
treatment 10 20 30 K
group 1 2 3 1 2 3 1 2 3 1 2 3
time 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
Your verbal description for the first one:
treatment K vs. treatment 10 for group 1 at time 3h
With the 3-factor interaction on the MODEL statement, 4 of terms contain
treatment, so need 1 and -1 at respective comparison levels of treatment
PROC MIXED;
MODEL hydration = treatment | group | time;
ESTIMATE 'k vs 10 | g=1, tm=2' treatment -1 0 0 1
treatment*group -1 0 0 0 0 0 0 0 0 1 0 0
treatment*time 0 -1 0 0 0 0 0 1
treatment*group*time 0 -1 0 0 0 0 0 0 0
0 0 0 0 0 0 0 0 0 0 1 0 0 0 0;
etc...
Given this statement works, the others are straightforward to edit by
examining the factors and placement of levels.
However, one of my new "favorite" topics is to demonstrate the LSMESTIMATE
statement from GLIMMIX which makes it much easier to compute contrasts of
the LSMEANS, first starting with your verbal descriptions:
Note, that GLIMMIX doesn't work with the type of REPEATED statement you
chose so this extension may not be relevant, but is an excellent lead-in
to showing how to write
non-positional parameter inputs in GLIMMIX (also available on the ESTIMATE
and CONTRAST statements)
Write out your descriptions of desired contrasts:
1. treatment K vs. treatment 10 for group 1 at time 3h
2. treatment K vs. treatment 20 for group 1 at time 3h
3. treatment K vs. treatment 30 for group 1 at time 3h
For "contrast" number 1., enter two sets of closed brackets and add the -1
and 1 coefficients in each followed by a comma and with placeholders for
all 3 variables:
[-1, . . . ] [1, . . . ]
The first contrast 1. states you want to compare level 4 with level 1 of
treatment (first variable)
[-1, 4 . .] [1, 1 . . ]
Enter the stated levels of the other two variables:
[-1, 4 1 2] [1, 1 1 2]
The LSMESTIMATE tatement allows for multiple contrasts, so easy to edit
the other two by changing treatment '1' to '2' and '3':
1. [-1, 4 1 2] [1, 1 1 2]
2. [-1, 4 1 2] [1, 2 1 2]
3. [-1, 4 1 2] [1, 3 1 2]
^^^
Then add all three sets into the LSMESTIMATE statement followed by a comma
and include the 3-factor interaction and descriptions, also add options cl
and adjust=....
PROC GLIMMIX;
MODEL hydration = treatment | group | time;
...
LSMESTIMATE treatment*group*time 'K vs 10 |grp=1, time 2' [1, 4 1 2] [-1,
1 1 2],
'K vs 20 |grp=1, time 2' [1, 4 1 2] [-1,
2 1 2],
'K vs 30 |grp=1, time 2' [1, 4 1 2] [-1,
3 1 2] / cl adjust=simulate(nsamp=40000 seed=121121);
run;
You'll see the output for this one statement gives the same results as
writing _3_ rather tedious ESTIMATE statements with all the 0's and 1's,
plus it
will also will compute multiple comparison adjustmens ... usually simulate
is a good choice here.
I've found many situations recently where familiarity with this coding
leads to great efficiency and ease in making the desired comparisons of
LSMEANS, esp, when you have multiple coefficients across factor levels and
they aren't just 1 and -1 and 0.
Robin High
UNMC
From:
Anne <anne.daubmann@GMX.DE>
To:
SAS-L@LISTSERV.UGA.EDU
Date:
06/08/2010 07:00 AM
Subject:
Using the CONTRAST Statement in PROC MIXED for 3 fixed effects
Sent by:
"SAS(r) Discussion" <SAS-L@LISTSERV.UGA.EDU>
Hello,
First of all I would like to introduce my study. I have a study population
divided into 3 different groups. For each group each subject got 4
different treatments to the skin at two different points in time. Now I
want to analyze whether the group, the treatment and the time has any
influence on the skin hydration.
The data have the following structure (the points in between shall show
that there are more subjects than two subjects in each group):
data t;
input subject Treatment$ group$ time$ hydration;
cards;
1 K 1 3h 0.424301601
1 10 1 3h 0.66230796
1 20 1 3h 0.935552903
1 30 1 3h 0.314342945
1 K 1 24h 0.193648359
1 10 1 24h 0.226547532
1 20 1 24h 0.555516816
1 30 1 24h 0.420854877
2 K 1 3h 0.894045221
2 10 1 3h 0.666443159
2 20 1 3h 0.984863413
2 30 1 3h 0.920030913
2 K 1 24h 0.144518192
2 10 1 24h 0.203838493
2 20 1 24h 0.638158456
2 30 1 24h 0.576927595
. . . . .
. . . . .
. . . . .
9 K 2 3h 0.924301601
9 10 2 3h 1.16230796
9 20 2 3h 1.435552903
9 30 2 3h 0.814342945
9 K 2 24h 0.693648359
9 10 2 24h 0.726547532
9 20 2 24h 1.055516816
9 30 2 24h 0.920854877
10 K 2 3h 1.394045221
10 10 2 3h 1.166443159
10 20 2 3h 1.484863413
10 30 2 3h 1.420030913
10 K 2 24h 0.644518192
10 10 2 24h 0.703838493
10 20 2 24h 1.138158456
10 30 2 24h 1.076927595
. . . . .
. . . . .
. . . . .
17 K 3 3h 1.024301601
17 10 3 3h 1.26230796
17 20 3 3h 1.535552903
17 30 3 3h 0.914342945
17 K 3 24h 0.793648359
17 10 3 24h 0.826547532
17 20 3 24h 1.155516816
17 30 3 24h 1.020854877
18 K 3 3h 1.494045221
18 10 3 3h 1.266443159
18 20 3 3h 1.584863413
18 30 3 3h 1.520030913
18 K 3 24h 0.744518192
18 10 3 24h 0.803838493
18 20 3 24h 1.238158456
18 30 3 24h 1.176927595
. . . . .
;
run;
To get the global effect I run the following repeated measure analysis of
variance:
proc mixed data = t;
class subject treatment group time;
model hydration = treatment group time;
repeated group time / subject = subject type = un@un;
random subject/ subject = subject;
run;
Now I am interested e.g. in the p-value of the differences in the means of
the pairwise comparisons, for example treatment K vs. treatment 10 for
group 1 at time 3h, treatment K vs. treatment 20 for group 1 at time 3h
and
treatment K vs. treatment 30 for group 1 at time 3h. As far as I know, I
have to take the CONTRAST statement but I have some problems to formulate
the contrast correctly. Can anyone help me with the contrasts?
Thanks a lot,
Anne
