3
$\begingroup$

I don't quite understand what is going on in the correlation coefficient formula. In the numerator we have the covariance, and in the denominator we have the standard deviation of variable x multiplied by the standard deviation of variable y.

So ultimately it is a ratio of covariance to the product of the two standard deviations.

What is dividing by the product of the two standard deviations doing to help us determine the corrrelation ?

I have tried to draw it out to help me visually understand it as I find this helps, but am stuck as to what I should be looking for.

$\endgroup$
1
$\begingroup$

Covariance and correlation coefficient measure essentially the same effect: How 'linked'[1,2] are two variables, i.e. if $X$ increases, how much will $Y$ increase on average?

The problem with covariance is that its value depends on the scales of the two variables: The value of the covariance doesn't tell you much if you don't also know over what range $X$ and $Y$ vary. Therefore, you don't know whether $X$ is more 'linked' with $Y$ or with $Z$ if you only know that, say, $cov(X,Y)=1000$ and $cov(X,Z)=0.1$. Maybe $Y$ is income in dollars (with big spreads), and $Z$ is percentage of time spent brushing teeth (very small spreads) -- then $X$ (amount spent on toothbrushes) may be more linked to $Z$, although the covariance value is quite lower [3].

To account for that, the correlation coefficient norms the covariance: We divide the covariance by the spreads (measured as standard deviations) of $X$ and $Y$. If you do the math (or run some simulations), you'll see that the correlation coefficient ranges from -1 (complete negative dependence) over 0 (no [linear] dependence) to 1 (complete positive dependence). Thus, it's possible to compare the degrees of 'linkage' between different pairs of variables. And after you've used them for a while, you get a feeling for how much 'linkage' exists for a certain correlation coefficient.


[1] I'd normally use 'correlated' instead of 'linked', but that might be confused with the correlation coefficient in this answer.

[2] To be more precise: 'linearly linked'. There are lots of examples where there is a clear relationship between $X$ and $Y$, but their correlation (and thus covariance) is zero, e.g. if the scatter plot of the two variables looks like a circle or a cross.

[3] And the covariance values will change if you change the units of the variables, e.g. if you express $Y$ in milliseconds or $Z$ in fraction of average income.

$\endgroup$
  • $\begingroup$ great I think I'm making more sense of it. Having looked at it further it kind of looks like a pair of Z-scores …?? (X- xbar)/SDx and (Y-ybar)/SDy……or is this not a good way to look at it ? $\endgroup$ – B.Miller Sep 3 '14 at 19:47
  • $\begingroup$ @B.Miller Yes, if you mean the similarity of normalizations: a z score normalizes the distance from mean (to the standard deviation); correlation coefficients similarly normalize the covariance (to the product of the two standard deviations). But don't try to see the correlation coefficient as a product of the Z scores of the two variables -- the Z scores are for each variable individually; the covariance and correlation coefficient explicitly describe the interaction between the variables. $\endgroup$ – dobiwan Sep 5 '14 at 11:33

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.