N/A
Posts: 0

# Contrast problem

Hello all,

I have a problem in estimating a contrast in a data set where there are 2 lines, 6 year-seasons (YS), 10 animals and 1 trait (TR) in 18 observations.

The desired contrast is X1-X2 where,
X1 = (mean of Line1+mean of Line1*YS) and
X2 = (mean of Line2+mean of Line2*YS).

I calculated the result manually:

X1 is manually estimated as:
Line1 mean = (66/10) = 6.6
Mean of Line1*YS1 = (28/5) = 5.60
Mean of Line1*YS2 = (30/4) = 7.50
Mean of Line1*YS3 = (8/1) = 8.00
Then X1 = 6.6 + 5.6 + 7.5 + 8 = 27.70

X2 is manually estimated as:
Line2 mean = (48/8) = 6.00
Mean of Line2*YS1 = (20/3) = 6.67
Mean of Line2*YS2 = (19/3) = 6.33
Mean of Line2*YS3 = (9/2) = 4.50
Then X2 = 6.00 + 6.67 + 6.33 + 4.50 = 23.5

X1-X2 = 27.7-23.5 = 4.2
*/
------------------------
and by SAS as:
------------------------
title " means";
data test;
input Line YS Animal TR @@;
cards;
1 1 1 5 1 1 1 6 1 1 2 5 1 2 2 8 1 2 3 7 1 2 3 9 1 3 3 8 1 1 4 7 1 2 4 6 1 1 5 5 2 1 6 6 2 2 6 5 2 3 7 4 2 1 7 9 2 2 8 8 2 1 8 5 2 2 9 6 2 3 10 5
;
run;

proc sort data=test; by ys line;
proc means noprint data=test;
class ys line;
var tr ;
output out=test2 mean= avg_tr;
run;
proc print data=test2;
run;

data line1 line2 ; set test2;
if line=1 then output line1;
if line=2 then output line2;
run;

proc means noprint data=line1;
var avg_tr; ;
output out=test3 sum= X1;
run;

proc means noprint data=line2;
var avg_tr; ;
output out=test4 sum= X2;
run;

data final;
merge test3 test4;
diff=x1-x2;
run;
title "The desired contrast X1-X2";
proc print data=final;
run;

-----------------------------------------------------------------------------------------------------------
The result is the same by both methods, but have two problems
1-I'd like to get the standard error of the difference between the X1 and X2.
2-I think that there is a duplicated intercept in the model. How to exclude it?

Is there any suggestions?
Posts: 3,805

## Re: Contrast problem

Using the estimateable functions for the LS means I came up with this. Notice that the LSmeans for LINE are different from the means give by PROC MEANS.

[pre]
data test;
input Line YS Animal TR @@;
cards;
1 1 1 5 1 1 1 6 1 1 2 5 1 2 2 8 1 2 3 7 1 2 3 9 1 3 3 8 1 1 4 7 1 2 4 6 1 1 5 5
2 1 6 6 2 2 6 5 2 3 7 4 2 1 7 9 2 2 8 8 2 1 8 5 2 2 9 6 2 3 10 5
;;;;
run;
proc print;
run;

proc mixed data=test;
class line ys;
model tr = line|ys;
lsmeans ys|line / e;
estimate 'Line 1 Mean' intercept 3 line 3 0 ys 1 1 1 line*ys 1 1 1 0 0 0 / divisor=3;
estimate 'Line 1*ys 1' intercept 3 line 3 0 ys 3 0 0 line*ys 3 0 0 0 0 0 / divisor=3;
estimate 'Line 1*ys 2' intercept 3 line 3 0 ys 0 3 0 line*ys 0 3 0 0 0 0 / divisor=3;
estimate 'Line 1*ys 3' intercept 3 line 3 0 ys 0 0 3 line*ys 0 0 3 0 0 0 / divisor=3;
Estimate 'x1' intercept 12 line 12 0 ys 4 4 4 line*ys 4 4 4 0 0 0 / divisor=3;
Estimate 'x2' intercept 12 line 0 12 ys 4 4 4 line*ys 0 0 0 4 4 4 / divisor=3;
Estimate 'x1-x2' intercept 0 line 12 -12 ys 0 0 0 line*ys 4 4 4 -4 -4 -4 / divisor=3;
run;
quit;
[/pre]
N/A
Posts: 0

## Re: Contrast problem

data_null_;
The final result is different from the manual one (4.8 vs 4.2). I think that it is due to different procs. LSmeans vs means.

Important!!
This data set is a small example. How to construct such contrasts in larger data set (say 200 levels of year-season and 400 interactions)?
I'm looking for a general and simpler way.
Posts: 3,805

## Re: Contrast problem

> data_null_;
> The final result is different from the manual one
> (4.8 vs 4.2). I think that it is due to different
> procs. LSmeans vs means.
Yes, the result is different if you had complete balance they would be the same. I would use the lsmeans.

> Important!!
> This data set is a small example. How to construct
> such contrasts in larger data set (say 200 levels of
> year-season and 400 interactions)?
> I'm looking for a general and simpler way.

Do you mean more levels or more factors?

I don't know enough to know the "proper" way to do this. Here is the brute force method. You should be able to generalize this to suite your actual data.

[pre]
data test;
input Line YS Animal TR @@;
cards;
1 1 1 5 1 1 1 6 1 1 2 5 1 2 2 8 1 2 3 7 1 2 3 9 1 3 3 8 1 1 4 7 1 2 4 6 1 1 5 5
2 1 6 6 2 2 6 5 2 3 7 4 2 1 7 9 2 2 8 8 2 1 8 5 2 2 9 6 2 3 10 5
;;;;
run;
proc print;
run;
ods listing close;
ods trace on;
proc mixed data=test;
class line ys;
model tr = line|ys;
lsmeans line line*ys / e;
ods output coef=Coef;
run;
quit;
ods trace off;
ods listing;
proc contents data=Coef varnum;
run;
proc print data=coef;
run;

filename FT55F001 temp;
data coef2;
set coef(where=(lmatrix eq 1) drop=row3-row6 rename=(Row1-Row2=L1-L2)) end=eof;
by notsorted Effect Line YS;
set coef(where=(lmatrix eq 2));
by notsorted Effect Line YS;
x1 = sum(of L1 Row1-Row3);
x2 = sum(of L2 Row4-Row6);
dif = x1-x2;

file FT55F001;
if _n_ eq 1 then put +3 'Estimate "x1-x2"';
if first.effect then put +6 effect @;
put dif @;
if last.effect then put;
if eof then put +6 ';';
format dif best32.;
run;
proc print;
run;
proc mixed data=test;
class line ys;
model tr = line|ys;
*lsmeans line line*ys;
%inc FT55F001 / source2;
run;
quit;
[/pre]
N/A
Posts: 0