When conducting research we are restricted in the number of variables that
can be examined at one time. For any particular sample, it may be that two
variables are very highly correlated and can not be distinguised from one
another numerically in a particular sample.  Such data is said to be
confounded.  There are only so many patterns to go around so some
confounding is bound to happen, especially when the variables have few
levels and small sample sizes. That is, when we do not look very deeply into
things. The issue of confounding has come up in discussions of corresponding
regressions lately and I thought I would throw out a method I use when
interpreting factor analyses and correlations.

Lets say we have two variables x1 and x2 that are highly correlated.  I try
to imagine all of the possible combinations of these two variables, whether
theyexist or not. The frequency of their occurance beyond even one
possibility does not much concern me. Let's say x1 is honesty and x2 is
professional status and that these two variables are highly correlated. If I
can think of any examples where dishonest people have high professional
status, then I realize the independence of these constructs. This would be a
test that all scientists should ask themselves when confronting the
complexity of phenomena.  If you can imagine exceptions to the correlation,
then you should acknowledge that the constructs could be sampled in such a
way that the high correlation is replaced by a zero correlation.

Let's take the example of mileage and engineering design in automobiles. If
the engineers know what they are doing and have the power to implement their
designs, then it is highly unlikely that they will build a model that does
not obtain the mileage standards desired. In this case, it is very difficult
to imagine a model of car that is manufactured to conserve x amount of
gasoline but that in fact does not so conserve.  For this reason, design or
intention is confounded with mileage.

Can you think of variables that could be confounded by this logical
orthogonalization method?  Do "professional" statisticians all do this
anyway?  Would such a method point the way to intentional sampling rather
than reliance on convenience samples?

Bill Chambers




.
.
=================================================================
Instructions for joining and leaving this list, remarks about the
problem of INAPPROPRIATE MESSAGES, and archives are available at:
.                  http://jse.stat.ncsu.edu/                    .
=================================================================

Reply via email to