Russel-Research-Method-in-Anthropology

Bivariate Analysis: Testing Relations 647
cant correlations in a symmetric matrix of 100 variables. If 50 of them (4,950/
100) might be the result of chance, then how can you decide which 50 they
are? You can’t. You can never know for sure whether any particular correlation
is the result of chance. You simply have to be careful in your interpretation of
every correlation in a matrix.
Use the shotgun. Be as cavalier as you can in looking for statistically significant
covariations, but be very conservative in interpreting their substantive
importance. Correlations are hints to you that something is going on between
two variables. Just keep in mind that the leap from correlation to cause is often
across a wide chasm.
If you look at table 20.18 again, you can see just how risky things can be.
A correlation of .60 is significant at the 1% level of confidence with a sample
as small as 30. Notice, however, that the correlation in the population is 99%
certain to fall between .20 and .83, which is a pretty wide spread. You
wouldn’t want to build too big a theory around a correlation that just might be
down around the .20 level, accounting for just 4% of the variance in what
you’re interested in!
Remember these rules:
1. Not all significant findings at the 5% level of confidence are equally important. A
very weak correlation of .10 in a sample of a million people would be statistically
significant, even if it were substantively trivial. By contrast, in small samples,
substantively important relations may show up as statistically insignificant.
2. Don’t settle for just one correlation that supports a pet theory; insist on several,
and be on the lookout for artifactual correlations.
Fifty years ago, before statistical packages were available, it was a real pain
to run any statistical tests. It made a lot of sense to think hard about which of
the thousands of possible tests one really wanted to run by hand on an adding
machine.
Computers have eliminated the drudge work in data analysis, but they
haven’t eliminated the need to think critically about your results. If anything,
computers—especially those little ones that fit in your backpack and have fullfeatured
statistics packages that do everything from 2 to factor analysis—
have made it more important than ever to be self-conscious about the interpretation
of statistical findings. But if you are self-conscious about this issue, and
dedicated to thinking critically about your data, then I believe you should take
full advantage of the power of the computer to produce a mountain of correlational
hints that you can follow up.
Finally, by all means, use your intuition in interpreting correlations; common
sense and your personal experience in research are powerful tools for