Xin,

Since your question raises a number of questions, it cannot be answered

with a simple, one line, answer.

First, since you have continuous variables, why make them binary and need

chi square? Keep them as they are and use pearson product-moment

correlations. That you can do, easily, with proc corr.

Second, you do realize that with only 20 subjects and over a thousand

correlations (or chi-squares for that matter), many of them will show what

appear as "real" correlation when, in fact, one would expect many of them

to look real through chance alone.

Thus, if you use .05 as your alpha, you would expect 50 of every 1,000

relationships to show a false significant correlation.

I'm sure that others can point out even more problems with what you are

attempting.

Art

---------