Rich Ulrich <[EMAIL PROTECTED]> wrote in 
news:[EMAIL PROTECTED]:

> How do *you*  describe the correlation of (say)  
> height with weight of a set of objects, where 
> every object is exactly the same height?
> 
>  - well, there is no "co"- variation, so you might
> settle for zero.  But I think that depends on how 
> limited your needs are.

It's rather odd that "intuitively" to me, it depends on whether or not you 
treat one variable as "independent" and another as "dependent."  If I look 
at a scatterplot where all the points are arrayed along a horizontal line, 
my first thought is "zero correlation" whereas if I see one where all the 
points fall on a vertical line, my first thought is "not enough information 
to tell whether there is a correlation."  But mathematically, that's 
nonsensical; zero correlation implies a roughly circular pattern of points 
(I like Stephen Jay Gould's heuristic that correlation measures the 
"skinniness" of a scatterplot), and it's rather obvious that if you simply 
interchange axes, there's no difference between the two situations.

I think this is a case where we're encountering what Sir Francis Bacon 
called an "idol," a prejudice of thought that distorts our view of reality.  
We want to believe that correlation is defined for any pair of random 
variables, but in fact it's meaningless if one of those RVs has all its 
density concentrated at a single point.  We somehow expect that because we 
call it a random "variable" it has to *vary*, but the definition of a 
random variable is merely "a function whose domain is a probability space" 
and constants meet that definition.
.
.
=================================================================
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