回复 第5楼 的 stella2011：The short story is that they are the same.
To ease your 1st concern, let's think of a simpler case where you could have used a paired t-test. For each subject, you have a pair of measurements. All you care about is the difference in means. As you suggested, you can fit a bivariate normal model with unstructured covariance matrix for the two observations within each single subject. Then, you can find the blue of the contrast of interest, i.e. the difference in means, and its stderr. Now what you get is the same thing as a paired t-test.
You can run the following SAS code to see the equivalence:
<br />
data d;<br />
call streaminit(2049);<br />
do sub = 1 to 10;<br />
b=rannor(2049)*3;<br />
do trt = 1 to 2;<br />
y=b+ rannor(2049)+trt;<br />
output;<br />
end;<br />
end;<br />
run;</p>
<p>proc mixed data=d;<br />
class trt sub;<br />
model y=trt / ddfm=kr;<br />
repeated / sub=sub type=un;<br />
ods select tests3;<br />
run;</p>
<p>data d2;<br />
call streaminit(2049);<br />
do sub=1 to 10;<br />
b=rannor(2049)*3;<br />
x1=b+rannor(2049)+1;<br />
x2=b+rannor(2049)+2;<br />
output;<br />
end;<br />
run;<br />
proc ttest data=d2;<br />
paired x1*x2;<br />
ods select ttests;<br />
run;<br />
</p>
What a multivariate model provides, in addition to those provided by the paired test, is the inference on covariances. Paired analysis assumes that you don't care that covariance. So there is really no need to bother ourselves in modeling it.
When is the paired design better then CRD? As I said earlier, it has better power if the correlation within each subject is positive, and worse if the correlation is negative. That's why I believe tedzzx's design is better.
Now back to the tedzzx's problem, CMH is the analog of the blocking/pairing for normal data. And the same principle applies -- randomizing, blocking, replicating.
Your 2nd concern about the goodness of asymptotic approximation in small samples can be solved by doing permutations. I will refer both exact and asymptotic tests as a CMH test.
